| hw_id,device_raw,device_count,backend,model,quant,params_b,size_gib,fa,ngl,threads,ubatch,depth,pp_n,pp_tps,tg_n,tg_tps,build,build_num,n_samples,date,notes,engine | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,12,512,0,512,1277.12,128,39.71,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 17.55; tg stddev 0.04",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,12,512,4096,512,1151.05,128,37.73,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 11.18; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,12,512,16384,512,950.78,128,36.79,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 6.81; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,12,512,0,512,598.35,128,41.12,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 16.28; tg stddev 0.16; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,12,512,0,512,476.43,128,34.64,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 10.00; tg stddev 0.71; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,12,512,4096,512,583.06,128,39.55,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 4.93; tg stddev 0.55; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,12,512,16384,512,557.1,128,38.67,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 5.92; tg stddev 0.28; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,12,512,0,512,476.61,128,45.66,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 7.91; tg stddev 0.20; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,12,512,0,512,365.02,128,35.31,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 6.48; tg stddev 0.21; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,12,512,4096,512,471.52,128,42.54,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 7.31; tg stddev 0.56; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,12,512,16384,512,400.83,128,37.82,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 1.58; tg stddev 0.16; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,12,512,0,512,1983.02,128,61.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 45.83; tg stddev 0.06",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,12,512,4096,512,1708.77,128,55.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 19.20; tg stddev 0.05",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,12,512,16384,512,1182.57,128,43.02,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 8.30; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,12,512,0,512,2086.63,128,39.25,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 42.87; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,12,512,4096,512,1776.4,128,36.59,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 15.71; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,12,512,16384,512,1215.46,128,30.64,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 7.60; tg stddev 0.02",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,12,512,0,512,2996.74,128,85.27,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 115.70; tg stddev 0.10",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,12,512,4096,512,2843.96,128,83.15,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 46.46; tg stddev 0.35",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,12,512,16384,512,2425.84,128,75.86,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 29.74; tg stddev 0.16",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,12,512,0,512,3133.42,128,62.6,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 107.71; tg stddev 0.12",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,12,512,4096,512,2992.45,128,61.27,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 49.17; tg stddev 0.10",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,12,512,16384,512,2506.95,128,57.05,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 40.31; tg stddev 0.08",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,12,512,0,512,1849.31,128,55.28,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 43.17; tg stddev 0.08",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,12,512,4096,512,1800.39,128,54.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 18.40; tg stddev 0.08",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,12,512,16384,512,1613.1,128,50.98,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 14.50; tg stddev 0.06",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,12,512,0,512,1945.31,128,36.6,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 47.53; tg stddev 0.04",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,12,512,4096,512,1891.49,128,36.12,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 18.52; tg stddev 0.04",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,12,512,16384,512,1683.59,128,34.62,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 16.59; tg stddev 0.03",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,12,512,0,512,360.8,128,41.13,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 8.85; tg stddev 0.35; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,12,512,4096,512,342.37,128,40.74,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.29; tg stddev 0.38; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,12,512,16384,512,328.8,128,39.76,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 2.04; tg stddev 0.61; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,12,512,0,512,1077.18,128,57.61,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 34.32; tg stddev 0.42; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,12,512,0,512,785.07,128,43.2,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 30.38; tg stddev 0.42; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,12,512,0,512,616.4,128,32.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 21.36; tg stddev 2.22; MoE expert offload: --n-cpu-moe 18 (expert tensors of the first 18 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,12,512,4096,512,1067.79,128,56.01,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 14.82; tg stddev 0.51; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060-12,NVIDIA GeForce RTX 3060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,12,512,16384,512,969.92,128,53.71,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen Threadripper 1920X 12-Core Processor, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 4.17; tg stddev 0.61; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-3060,NVIDIA GeForce RTX 3060 12GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,52,512,0,4096,232.87,128,31.97,3e4dfa8f3,11658,2,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, vast.ai, CPU Intel Xeon Platinum 8272L, RAM 1486.9 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 2; Strata comparison run; MoE expert offload --n-cpu-moe 22; excluded from site medians",llama.cpp | |
| rtx-3060,NVIDIA GeForce RTX 3060 12GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,52,512,32768,4096,193.17,128,27.16,3e4dfa8f3,11658,1,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, vast.ai, CPU Intel Xeon Platinum 8272L, RAM 1486.9 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 1; Strata comparison run; MoE expert offload --n-cpu-moe 26 (lowest that fits at this depth); excluded from site medians",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,24,512,0,512,3271.97,128,86.54,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 87.38; tg stddev 0.16",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,24,512,4096,512,2960.69,128,82.7,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 49.26; tg stddev 0.15",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,24,512,16384,512,2438.93,128,80.19,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 30.26; tg stddev 0.11",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,24,512,0,512,3408.05,128,57.24,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 80.86; tg stddev 0.06",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,24,512,4096,512,3080.06,128,55.48,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 50.13; tg stddev 0.08",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,24,512,16384,512,2523.11,128,54.32,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 42.38; tg stddev 0.07",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,24,512,0,512,4663.86,128,140.26,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 185.91; tg stddev 0.50",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,24,512,4096,512,4285.15,128,132.74,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 106.77; tg stddev 1.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,24,512,16384,512,3667.99,128,127.42,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 83.73; tg stddev 0.66",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,24,512,0,512,1292.43,128,38.78,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 11.98; tg stddev 0.01",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,24,512,4096,512,1182.19,128,36.8,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 8.98; tg stddev 0.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,24,512,16384,512,968.76,128,35.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 6.11; tg stddev 0.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,24,512,0,512,1897.06,128,53.15,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 24.13; tg stddev 0.01",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,24,512,4096,512,1793.83,128,51.07,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 12.58; tg stddev 0.04",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,24,512,16384,512,1473.55,128,45.91,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 9.27; tg stddev 0.01",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,24,512,0,512,4304.46,128,206.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 200.58; tg stddev 1.40",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,24,512,4096,512,3709.85,128,188.55,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 59.15; tg stddev 1.59",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,24,512,16384,512,2381.21,128,149.37,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 33.00; tg stddev 0.79",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,24,512,0,512,4898.11,128,134.6,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 159.87; tg stddev 0.44",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,24,512,4096,512,4285.15,128,123.07,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 84.02; tg stddev 0.15",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,24,512,16384,512,2940.27,128,98.69,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 32.32; tg stddev 0.21",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,24,512,0,512,5109.2,128,91.35,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 134.56; tg stddev 0.09",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,24,512,4096,512,4418.49,128,85.86,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 95.27; tg stddev 0.14",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,24,512,16384,512,2997.95,128,73.05,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 43.61; tg stddev 0.08",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,24,512,0,512,7365.78,128,179.09,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 449.72; tg stddev 1.52",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,24,512,4096,512,7162.07,128,175.92,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 250.20; tg stddev 0.37",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,24,512,16384,512,5994.92,128,163.16,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 147.61; tg stddev 0.55",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,24,512,0,512,7630.74,128,137.98,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 499.34; tg stddev 0.15",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,24,512,4096,512,7371.12,128,135.64,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 232.40; tg stddev 0.27",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,24,512,16384,512,6231.5,128,127.55,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 184.47; tg stddev 0.34",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,24,512,0,512,4632.1,128,122.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 160.82; tg stddev 0.16",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,24,512,4096,512,4544.73,128,120.15,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 83.25; tg stddev 0.09",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,24,512,16384,512,4027.01,128,113.85,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 65.33; tg stddev 0.14",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,24,512,0,512,4788.32,128,85.66,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 169.58; tg stddev 0.10",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,24,512,4096,512,4718.64,128,84.76,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 82.15; tg stddev 0.14",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,24,512,16384,512,4216.81,128,81.48,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 80.41; tg stddev 0.12",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,24,512,0,512,1515.77,128,42.18,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 19.65; tg stddev 0.02",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,24,512,4096,512,1462.74,128,41.69,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 15.01; tg stddev 0.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,24,512,16384,512,1276.35,128,40.1,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 16.82; tg stddev 0.06",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,24,512,0,512,4053.33,128,158.97,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 276.66; tg stddev 1.31",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,24,512,4096,512,3868.21,128,157.6,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 122.27; tg stddev 1.57",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,24,512,16384,512,3430.0,128,150.59,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 98.77; tg stddev 1.41",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,24,512,0,512,5857.32,128,214.34,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 282.22; tg stddev 1.51",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,24,512,4096,512,5647.65,128,208.37,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 94.54; tg stddev 1.54",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,24,512,16384,512,4546.31,128,193.84,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 95.03; tg stddev 0.54",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,24,512,0,512,5951.86,128,204.22,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 252.23; tg stddev 1.07",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,24,512,4096,512,5696.08,128,198.28,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 97.20; tg stddev 1.15",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,24,512,16384,512,4521.59,128,185.17,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.126.09, vast.ai, CPU AMD Ryzen Threadripper 3960X 24-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 96.17; tg stddev 0.73",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,96,512,32768,512,1931.21,128,78.39,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 57.76; tg stddev 0.26",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,96,512,65536,512,1407.72,128,74.7,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 25.32; tg stddev 0.20",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,96,512,131072,512,904.93,128,67.76,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 10.52; tg stddev 0.16",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,96,512,32768,512,1188.33,128,40.65,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 17.57; tg stddev 0.01",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,96,512,32768,512,1609.47,128,116.4,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 48.95; tg stddev 0.56",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,96,512,32768,512,2014.44,128,77.68,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 45.34; tg stddev 0.17",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,96,512,65536,512,1284.16,128,54.47,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 13.11; tg stddev 0.04",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,96,512,131072,512,739.68,128,34.06,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 6.60; tg stddev 0.07",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,96,512,32768,512,4576.75,128,146.97,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 257.83; tg stddev 1.02",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,96,512,65536,512,3454.06,128,123.77,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 118.07; tg stddev 0.39",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,96,512,131072,512,2254.38,128,93.97,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 85.81; tg stddev 0.27",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,96,512,32768,512,3413.32,128,106.21,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 146.13; tg stddev 0.25",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,96,512,65536,512,2740.12,128,93.79,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 51.28; tg stddev 0.25",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,96,512,131072,512,1903.9,128,75.84,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 47.53; tg stddev 0.23",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,96,512,32768,512,1081.03,128,38.48,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 26.66; tg stddev 0.06",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,96,512,65536,512,836.83,128,35.15,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 22.49; tg stddev 0.07",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,96,512,32768,512,3465.26,128,177.53,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 160.09; tg stddev 1.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,96,512,65536,512,2491.49,128,151.98,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 56.72; tg stddev 0.35",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,96,512,131072,512,1574.52,128,116.95,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7642 48-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 16.59; tg stddev 0.27",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,32,512,0,512,1506.69,128,42.9,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 48.84; tg stddev 0.16",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,32,512,4096,512,1437.61,128,42.4,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 36.22; tg stddev 0.02",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,32,512,16384,512,1234.06,128,40.6,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 25.87; tg stddev 0.12",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,32,512,32768,512,1057.59,128,38.73,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 12.18; tg stddev 0.06",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,32,512,65536,512,820.06,128,35.48,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 11.13; tg stddev 0.02",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,32,512,0,512,1360.18,128,33.94,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 42.60; tg stddev 0.04",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,32,512,4096,512,1298.81,128,33.51,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 22.68; tg stddev 0.05",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,32,512,16384,512,1138.56,128,32.46,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 20.01; tg stddev 0.04",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,32,512,32768,512,984.28,128,31.24,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 615.71.10, vast.ai, CPU AMD EPYC 7452 32-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 11.21; tg stddev 0.03",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,128,512,0,4096,3637.96,128,148.17,3e4dfa8f3,11658,3,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.58.03, vast.ai, CPU AMD EPYC 7713, RAM 241.8 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 3; Strata comparison run",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,128,512,32768,4096,2716.12,128,132.84,3e4dfa8f3,11658,3,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.58.03, vast.ai, CPU AMD EPYC 7713, RAM 241.8 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 3; Strata comparison run",llama.cpp | |
| rtx-3090,NVIDIA GeForce RTX 3090,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,128,512,0,4096,730.32,128,23.08,3e4dfa8f3,11658,3,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.58.03, vast.ai, CPU AMD EPYC 7713, RAM 241.8 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 3; Strata comparison run; MoE expert offload --n-cpu-moe 20; excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,0,512,646.68,128,39.13,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 12.19; tg stddev 0.21; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,4096,512,640.67,128,35.79,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 2.26; tg stddev 0.20; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,16384,512,615.31,128,34.28,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 6.02; tg stddev 0.18; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,0,512,510.66,128,45.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 6.13; tg stddev 0.12; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,4096,512,508.68,128,40.11,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 5.13; tg stddev 0.45; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,16384,512,442.75,128,35.52,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 3.22; tg stddev 0.57; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,0,512,2444.75,128,51.27,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 25.80; tg stddev 0.06",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,4096,512,2093.54,128,45.59,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 18.11; tg stddev 0.04",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,16384,512,1376.49,128,34.39,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 6.15; tg stddev 0.02",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,0,512,3695.08,128,77.36,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 228.60; tg stddev 0.18",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,4096,512,3621.8,128,74.29,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 80.90; tg stddev 0.17",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,16384,512,3071.05,128,66.44,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 31.28; tg stddev 0.15",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,0,512,3948.31,128,51.32,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 111.14; tg stddev 0.05",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,4096,512,3756.23,128,49.96,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 53.30; tg stddev 0.07",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,16384,512,3191.8,128,46.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 35.72; tg stddev 0.05",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,0,512,2283.04,128,45.62,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 28.52; tg stddev 0.04",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,4096,512,2213.22,128,44.54,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 3.23; tg stddev 0.06",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,16384,512,1996.6,128,41.59,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 7.89; tg stddev 0.05",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,0,512,492.62,128,48.76,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 10.35; tg stddev 0.20; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,4096,512,466.88,128,47.91,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.73; tg stddev 0.22; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,16384,512,448.21,128,45.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.02; tg stddev 0.31; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,1057.96,128,49.79,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 33.78; tg stddev 0.34; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,844.79,128,39.37,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 27.28; tg stddev 1.26; MoE expert offload: --n-cpu-moe 18 (expert tensors of the first 18 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,4096,512,1042.12,128,48.72,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 10.29; tg stddev 0.11; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060,NVIDIA GeForce RTX 4060,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,16384,512,929.35,128,45.44,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.58.03, vast.ai, CPU AMD Ryzen 9 7945HX with Radeon Graphics, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 6.85; tg stddev 0.99; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,8,512,0,512,2185.72,128,35.43,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 36.58; tg stddev 0.03",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,8,512,4096,512,1983.13,128,33.5,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 14.28; tg stddev 0.03",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,8,512,16384,512,1660.02,128,32.68,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 12.36; tg stddev 0.02",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,8,512,0,512,2215.47,128,20.58,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 30.01; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,8,512,4096,512,2008.09,128,19.91,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 13.66; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,8,512,16384,512,1675.93,128,19.62,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 10.28; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,8,512,0,512,1039.71,128,44.66,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 26.18; tg stddev 0.05; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,8,512,0,512,684.65,128,32.9,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 17.49; tg stddev 0.30; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,8,512,0,512,525.43,128,26.67,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 13.17; tg stddev 0.02; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,8,512,4096,512,1015.16,128,42.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 6.63; tg stddev 0.07; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,8,512,16384,512,891.21,128,40.76,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 20.95; tg stddev 0.07; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,8,512,0,512,1205.69,128,19.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 0.60; tg stddev 0.00",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,8,512,4096,512,1132.12,128,18.18,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 5.90; tg stddev 0.00",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,8,512,0,512,873.6,128,54.2,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 7.10; tg stddev 0.22; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,8,512,0,512,541.74,128,35.58,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 4.05; tg stddev 0.63; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,8,512,0,512,404.91,128,27.01,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 3.24; tg stddev 0.04; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,8,512,4096,512,833.43,128,49.83,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 10.54; tg stddev 0.04; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,8,512,16384,512,706.04,128,42.02,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 3.13; tg stddev 0.07; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,8,512,0,512,3406.37,128,54.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 63.26; tg stddev 0.03",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,8,512,4096,512,2922.41,128,48.13,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 21.93; tg stddev 0.04",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,8,512,16384,512,1651.44,128,36.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 10.62; tg stddev 0.02",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,8,512,0,512,3558.94,128,32.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 63.60; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,8,512,4096,512,3035.26,128,30.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 21.92; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,8,512,16384,512,1657.39,128,25.05,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 11.80; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,8,512,0,512,5288.39,128,82.35,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 279.94; tg stddev 0.11",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,8,512,4096,512,5048.15,128,79.22,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 74.94; tg stddev 0.11",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,8,512,16384,512,3628.23,128,70.77,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 35.00; tg stddev 0.11",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,8,512,0,512,5330.27,128,54.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 269.16; tg stddev 0.04",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,8,512,4096,512,5116.16,128,52.92,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 54.54; tg stddev 0.05",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,8,512,16384,512,3669.46,128,49.03,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 30.56; tg stddev 0.04",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,8,512,0,512,3190.98,128,48.3,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 89.54; tg stddev 0.02",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,8,512,4096,512,3091.63,128,47.17,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 23.23; tg stddev 0.04",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,8,512,16384,512,2533.1,128,44.05,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 19.58; tg stddev 0.03",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,8,512,0,512,3359.51,128,30.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 85.29; tg stddev 0.01",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,8,512,4096,512,3249.15,128,29.94,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 31.37; tg stddev 0.02",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,8,512,16384,512,2610.15,128,28.65,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 5.14; tg stddev 0.02",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,8,512,0,512,549.76,128,42.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 12.40; tg stddev 0.04; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,8,512,0,512,401.28,128,33.97,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 8.50; tg stddev 0.32; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,8,512,4096,512,519.56,128,41.4,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.38; tg stddev 0.07; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,8,512,16384,512,499.85,128,39.96,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 4.26; tg stddev 0.06; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,8,512,0,512,4049.11,128,103.72,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 192.50; tg stddev 0.27",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,8,512,4096,512,3848.87,128,98.32,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 47.32; tg stddev 0.21",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,8,512,16384,512,3188.99,128,89.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 43.43; tg stddev 0.25",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,8,512,0,512,4109.9,128,94.43,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 116.45; tg stddev 0.13",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,8,512,4096,512,3861.23,128,89.8,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 22.79; tg stddev 0.14",llama.cpp | |
| rtx-4060ti-16,NVIDIA GeForce RTX 4060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,8,512,16384,512,3174.48,128,82.31,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz, RAM 62 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 42.00; tg stddev 0.22",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,0,512,2928.16,128,58.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 102.61; tg stddev 0.10",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,4096,512,2665.23,128,55.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 47.61; tg stddev 0.11",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,16384,512,2257.77,128,54.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 32.06; tg stddev 0.10",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,0,512,940.36,128,49.93,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 27.27; tg stddev 0.34; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,0,512,720.04,128,37.46,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 16.44; tg stddev 0.07; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,4096,512,924.65,128,48.45,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 8.22; tg stddev 0.11; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,16384,512,880.11,128,47.32,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 9.25; tg stddev 0.10; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,0,512,743.85,128,49.46,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 10.81; tg stddev 0.20; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,0,512,554.76,128,36.54,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 10.65; tg stddev 0.12; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,4096,512,740.76,128,44.42,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 13.73; tg stddev 0.18; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,16384,512,656.48,128,43.38,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 1.13; tg stddev 0.17; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,0,512,4489.19,128,89.69,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 150.07; tg stddev 0.17",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,4096,512,3879.76,128,78.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 71.08; tg stddev 0.20",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,16384,512,2303.0,128,57.9,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 16.22; tg stddev 0.08",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,0,512,4713.61,128,54.99,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 188.55; tg stddev 0.06",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,4096,512,4053.16,128,50.46,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 49.28; tg stddev 0.05",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,16384,512,2339.27,128,41.22,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 26.01; tg stddev 0.01",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,0,512,6623.36,128,129.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 597.02; tg stddev 0.58",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,4096,512,6458.65,128,124.37,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 222.05; tg stddev 0.53",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,16384,512,5202.55,128,110.34,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 146.00; tg stddev 0.44",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,0,512,7082.18,128,88.97,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 610.33; tg stddev 0.18",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,4096,512,6776.28,128,86.41,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 197.88; tg stddev 0.23",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,16384,512,5390.83,128,79.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 120.91; tg stddev 0.16",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,0,512,4236.74,128,79.8,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 199.17; tg stddev 0.16",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,4096,512,4108.63,128,77.75,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 67.58; tg stddev 0.16",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,16384,512,3557.34,128,72.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 35.61; tg stddev 0.16",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,0,512,4417.49,128,51.26,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 304.63; tg stddev 0.05",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,4096,512,4334.29,128,50.41,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 68.17; tg stddev 0.06",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,16384,512,3727.6,128,47.96,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 55.53; tg stddev 0.06",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,0,512,536.57,128,45.97,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 13.57; tg stddev 1.01; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,4096,512,510.43,128,46.4,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.17; tg stddev 0.15; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,16384,512,492.78,128,45.11,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.32; tg stddev 0.18; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,1844.86,128,72.46,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 57.81; tg stddev 0.09; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,1217.44,128,45.15,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 24.29; tg stddev 0.08; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,956.56,128,33.6,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 32.84; tg stddev 0.02; MoE expert offload: --n-cpu-moe 18 (expert tensors of the first 18 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,4096,512,1822.07,128,71.59,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 27.88; tg stddev 0.24; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4070,NVIDIA GeForce RTX 4070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,16384,512,1700.39,128,69.76,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 9 3950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 7.69; tg stddev 0.16; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,6,512,0,512,6465.41,128,101.95,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 33.96; tg stddev 0.08",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,6,512,4096,512,6357.26,128,97.75,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 510.46; tg stddev 0.11",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,6,512,16384,512,5290.2,128,94.7,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 92.58; tg stddev 0.19",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,6,512,0,512,6420.34,128,64.22,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 143.55; tg stddev 0.09",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,6,512,4096,512,6534.51,128,62.49,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 31.97; tg stddev 0.06",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,6,512,16384,512,5369.52,128,61.24,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 434.83; tg stddev 0.09",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,6,512,0,512,9164.92,128,169.27,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 586.17; tg stddev 1.61",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,6,512,4096,512,8915.07,128,160.61,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 356.89; tg stddev 0.93",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,6,512,16384,512,7815.94,128,152.57,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 171.92; tg stddev 0.94",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,6,512,0,512,2403.21,128,44.42,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 20.91; tg stddev 0.04",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,6,512,4096,512,2285.0,128,41.83,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 33.83; tg stddev 0.03",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,6,512,16384,512,1959.85,128,40.07,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 24.66; tg stddev 0.03",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,6,512,0,512,3302.93,128,61.56,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 16.05; tg stddev 0.07",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,6,512,4096,512,3403.7,128,58.74,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 141.78; tg stddev 0.13",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,6,512,16384,512,2790.83,128,51.83,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 85.55; tg stddev 0.04",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,6,512,0,512,8590.1,128,265.96,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 390.20; tg stddev 2.53",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,6,512,4096,512,7825.1,128,223.5,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 359.31; tg stddev 0.28",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,6,512,16384,512,5463.34,128,177.61,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 491.35; tg stddev 1.85",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,6,512,0,512,10766.36,128,158.52,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 675.66; tg stddev 0.76",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,6,512,4096,512,9685.92,128,143.06,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 457.48; tg stddev 0.21",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,6,512,16384,512,6688.8,128,111.62,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 808.77; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.08",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,6,512,0,512,10242.31,128,102.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 640.01; tg stddev 0.02",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,6,512,4096,512,10120.61,128,95.66,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 476.86; tg stddev 0.13",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,6,512,16384,512,6872.4,128,80.51,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 364.97; tg stddev 0.17",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,6,512,0,512,14882.2,128,220.18,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1299.90; tg stddev 0.68",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,6,512,4096,512,14713.29,128,213.33,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 665.75; tg stddev 0.66",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,6,512,16384,512,13052.24,128,194.56,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 640.84; tg stddev 0.86",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,6,512,0,512,15454.65,128,158.31,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 1163.16; tg stddev 0.29",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,6,512,4096,512,15490.0,128,154.99,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 805.05; tg stddev 0.17",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,6,512,16384,512,13436.15,128,144.69,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 461.98; tg stddev 0.34",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,6,512,0,512,9761.75,128,141.55,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 428.50; tg stddev 0.58",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,6,512,4096,512,10049.99,128,138.77,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 819.85; tg stddev 0.17",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,6,512,16384,512,9525.47,128,130.79,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 217.66; tg stddev 0.09",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,6,512,0,512,9849.56,128,95.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 59.74; tg stddev 0.13",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,6,512,4096,512,10238.08,128,93.77,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 780.70; tg stddev 0.12",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,6,512,16384,512,8886.98,128,90.03,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 658.09; tg stddev 0.20",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,6,512,0,512,2795.39,128,48.38,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 68.64; tg stddev 0.04",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,6,512,4096,512,2752.44,128,47.71,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 61.47; tg stddev 0.05",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,6,512,16384,512,2542.78,128,45.9,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 96.67; tg stddev 0.03",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,6,512,0,512,7962.13,128,190.49,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 657.50; tg stddev 1.18",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,6,512,4096,512,7966.81,128,187.06,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 569.86; tg stddev 2.18",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,6,512,16384,512,7198.28,128,177.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 489.61; tg stddev 2.03",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,6,512,0,512,11344.42,128,277.95,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 734.68; tg stddev 1.50",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,6,512,4096,512,11298.4,128,263.1,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 538.69; tg stddev 2.73",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,6,512,16384,512,9422.29,128,242.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1084.90; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.07",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,6,512,0,512,12257.18,128,258.4,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 616.82; tg stddev 1.92",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,6,512,4096,512,11510.77,128,245.77,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 550.44; tg stddev 1.55",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,6,512,16384,512,9622.1,128,227.89,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD Ryzen 5 7600X 6-Core Processor, RAM 30 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 536.00; tg stddev 1.53",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,32768,512,4507.11,128,92.94,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 250.10; tg stddev 0.32",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,65536,512,3359.03,128,87.57,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 61.88; tg stddev 0.17",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,131072,512,2239.75,128,78.97,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 24.99; tg stddev 0.20",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,16,512,32768,512,2556.65,128,45.83,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 50.79; tg stddev 0.03",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,32768,512,3821.42,128,133.55,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 223.50; tg stddev 1.31",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,32768,512,4899.02,128,87.51,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 46.95; tg stddev 0.17",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,65536,512,3107.95,128,60.36,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 72.63; tg stddev 0.09",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,131072,512,1770.24,128,37.17,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 32.84; tg stddev 0.01",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,32768,512,10398.99,128,175.22,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 838.41; tg stddev 1.55",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,65536,512,8099.92,128,145.21,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 306.59; tg stddev 0.68",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,131072,512,5184.48,128,107.41,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 116.08; tg stddev 0.35",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,32768,512,7955.97,128,122.02,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 397.51; tg stddev 0.21",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,65536,512,6347.38,128,106.53,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 157.79; tg stddev 0.46",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,131072,512,4522.07,128,85.0,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 110.21; tg stddev 0.25",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,16,512,32768,512,2334.5,128,44.06,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 97.75; tg stddev 0.08",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,16,512,65536,512,1867.31,128,39.95,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 62.68; tg stddev 0.07",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,32768,512,7891.5,128,217.26,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 318.95; tg stddev 2.27",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,65536,512,5816.0,128,181.07,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 143.46; tg stddev 0.31",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,131072,512,3663.03,128,136.72,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 50.76; tg stddev 0.71",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,6,512,0,512,3417.14,128,50.71,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 241.48; tg stddev 0.16",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,6,512,4096,512,3334.26,128,50.09,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 109.54; tg stddev 0.12",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,6,512,16384,512,2962.3,128,48.03,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 80.82; tg stddev 0.10",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,6,512,32768,512,2578.29,128,45.49,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 67.01; tg stddev 0.09",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,6,512,65536,512,2061.0,128,41.17,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 42.20; tg stddev 0.08",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,6,512,0,512,2995.65,128,39.72,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 176.32; tg stddev 0.08",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,6,512,4096,512,2912.47,128,39.31,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 75.07; tg stddev 0.07",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,6,512,16384,512,2626.11,128,38.02,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 76.53; tg stddev 0.07",llama.cpp | |
| rtx-4090,NVIDIA GeForce RTX 4090,1,CUDA,Qwen3.8 27B,Q6_K,27.78,20.46,1,99,6,512,32768,512,2322.46,128,36.4,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 580.159.03, vast.ai, CPU AMD Ryzen 5 3600 6-Core Processor, RAM 62 GiB; llama-bench model_type 'qwen35 27B Q6_K'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q6_K.gguf; pp stddev 46.34; tg stddev 0.06",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,128,512,0,512,1981.01,128,49.05,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 84.45; tg stddev 0.10",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,128,512,4096,512,1786.13,128,46.31,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 65.16; tg stddev 0.08",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,128,512,16384,512,1470.92,128,45.54,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 59.84; tg stddev 0.08",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,128,512,0,512,2054.86,128,30.99,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 83.07; tg stddev 0.04",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,128,512,4096,512,1862.11,128,29.88,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 52.65; tg stddev 0.04",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,128,512,16384,512,1536.85,128,29.56,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 32.74; tg stddev 0.03",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,128,512,0,512,934.15,128,12.35,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 33.75; tg stddev 0.44; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,128,512,0,512,643.21,128,7.56,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 19.37; tg stddev 0.14; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,128,512,0,512,484.92,128,5.45,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 9.32; tg stddev 0.15; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,128,512,4096,512,929.29,128,12.63,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 15.90; tg stddev 1.42; HIGH VARIANCE in tg (>10% of mean); MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,128,512,16384,512,903.34,128,12.57,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 13.33; tg stddev 0.54; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,128,512,0,512,1103.93,128,27.88,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 17.64; tg stddev 0.01",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,128,512,4096,512,1033.8,128,26.52,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 15.11; tg stddev 0.04",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,128,512,0,512,786.26,128,10.35,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 22.45; tg stddev 0.67; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,128,512,0,512,511.19,128,5.89,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 12.30; tg stddev 0.22; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,128,512,0,512,367.39,128,4.18,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 9.72; tg stddev 0.04; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,128,512,4096,512,758.72,128,9.57,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 15.11; tg stddev 0.14; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,128,512,16384,512,653.44,128,9.44,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 5.57; tg stddev 0.24; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,128,512,0,512,3060.55,128,77.69,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 141.63; tg stddev 0.13",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,128,512,4096,512,2613.43,128,69.07,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 87.17; tg stddev 0.30",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,128,512,16384,512,1422.97,128,53.05,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 44.11; tg stddev 0.09",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,128,512,0,512,3189.44,128,49.09,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 142.75; tg stddev 0.08",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,128,512,4096,512,2719.45,128,45.52,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 72.12; tg stddev 0.06",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,128,512,16384,512,1471.52,128,37.95,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 22.43; tg stddev 0.04",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,128,512,0,512,4775.88,128,107.64,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 365.99; tg stddev 0.21",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,128,512,4096,512,4500.46,128,103.26,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 259.86; tg stddev 0.28",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,128,512,16384,512,3838.55,128,94.48,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 177.76; tg stddev 0.24",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,128,512,0,512,4979.83,128,79.64,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 435.84; tg stddev 0.21",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,128,512,4096,512,4717.88,128,77.02,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 228.24; tg stddev 0.13",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,128,512,16384,512,3963.36,128,72.07,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 173.61; tg stddev 0.22",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,128,512,0,512,2955.05,128,70.02,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 137.37; tg stddev 0.09",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,128,512,4096,512,2857.91,128,68.02,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 99.10; tg stddev 0.22",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,128,512,16384,512,2581.77,128,64.14,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 67.85; tg stddev 0.11",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,128,512,0,512,3078.17,128,45.91,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 157.36; tg stddev 0.02",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,128,512,4096,512,2974.84,128,45.02,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 97.93; tg stddev 0.08",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,128,512,16384,512,2656.74,128,43.31,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 91.03; tg stddev 0.08",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,128,512,0,512,502.05,128,7.55,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 18.01; tg stddev 0.41; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,128,512,0,512,369.32,128,5.22,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 11.35; tg stddev 0.22; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,128,512,4096,512,481.62,128,7.45,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 5.14; tg stddev 0.35; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,128,512,16384,512,466.14,128,6.69,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 4.21; tg stddev 0.09; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,128,512,0,512,5143.56,128,134.09,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 397.44; tg stddev 0.44",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,128,512,4096,512,4704.89,128,124.98,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 211.27; tg stddev 0.48",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,128,512,16384,512,3673.85,128,117.9,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 137.91; tg stddev 0.97",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,128,512,0,512,5236.97,128,125.31,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 384.10; tg stddev 0.52",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,128,512,4096,512,4746.58,128,117.31,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 121.79; tg stddev 0.31",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,128,512,16384,512,3680.46,128,111.32,3e4dfa8f3,11658,3,2026-10-21,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7V12 64-Core Processor, RAM 251 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 131.77; tg stddev 0.86",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,34.66,20.81,1,99,36,512,0,4096,473.01,128,28.81,3e4dfa8f3,11658,2,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.84, vast.ai, CPU Intel Xeon E5-2695 v4, RAM 503.0 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 2; Strata comparison run; host load average 11-51 during the llama.cpp stages and 12-20 during the Strata stages, vs 90-150 on the first run; MoE expert offload --n-cpu-moe 13; excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,34.66,20.81,1,99,36,512,4096,4096,405.61,128,27.27,3e4dfa8f3,11658,2,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.84, vast.ai, CPU Intel Xeon E5-2695 v4, RAM 503.0 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 2; Strata comparison run; host load average 11-51 during the llama.cpp stages and 12-20 during the Strata stages, vs 90-150 on the first run; MoE expert offload --n-cpu-moe 13; excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,34.66,20.81,1,99,36,512,32768,4096,354.05,128,19.27,3e4dfa8f3,11658,1,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, driver 595.84, vast.ai, CPU Intel Xeon E5-2695 v4, RAM 503.0 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 1; Strata comparison run; host load average 11-51 during the llama.cpp stages and 12-20 during the Strata stages, vs 90-150 on the first run; MoE expert offload --n-cpu-moe 15 (lowest that fits at this depth); excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,34.66,20.81,1,99,36,512,0,130,108.47,768,31.51,3e4dfa8f3,11658,3,2026-10-28,"own run, llama-server from llama.cpp b11658 (commit 3e4dfa8f3), -fa on -ngl 99 -c 16384 --jinja, 3 realistic ~100-token chat prompts with 768-token replies (same as Strata), median of server timings; vast.ai, CPU Intel Xeon E5-2695 v4; Strata comparison run; llama-server chat; host load average 11-51 during the llama.cpp stages and 12-20 during the Strata stages, vs 90-150 on the first run; MoE expert offload --n-cpu-moe 13; excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,56,512,0,512,234.71,64,14.15,3e4dfa8f3,11658,1,2026-10-28,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), engine llama.cpp, vast.ai, CPU Intel Xeon Platinum 8273CL, RAM 761.0 GiB; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf; -fa 1 -ngl 99 -r 1; Strata comparison run; busy shared host (load average 90-150 from other tenants during the run), which mostly slows CPU-side expert work; MoE expert offload --n-cpu-moe 14; excluded from site medians",llama.cpp | |
| rtx-5060ti-16,NVIDIA GeForce RTX 5060 Ti 16GB,1,CUDA,Qwen3.6 35B-A3B,UD-Q4_K_XL,35.51,21.27,1,99,56,512,0,130,49.33,768,11.19,3e4dfa8f3,11658,3,2026-10-28,"own run, llama-server from llama.cpp b11658 (commit 3e4dfa8f3), -fa on -ngl 99 -c 16384 --jinja, 3 realistic ~100-token chat prompts with 768-token replies (same as Strata), median of server timings; vast.ai, CPU Intel Xeon Platinum 8273CL; Strata comparison run; llama-server chat; busy shared host (load average 90-150 from other tenants during the run), which mostly slows CPU-side expert work; MoE expert offload --n-cpu-moe 14; excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,0,512,3167.56,128,70.13,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 117.31; tg stddev 0.02",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,4096,512,2911.59,128,66.32,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 66.35; tg stddev 0.04",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,16,512,16384,512,2458.89,128,65.4,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 45.46; tg stddev 0.02",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,0,512,1220.85,128,32.32,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 45.47; tg stddev 0.22; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,0,512,959.86,128,26.56,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 29.31; tg stddev 0.26; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,4096,512,1206.57,128,30.99,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 15.85; tg stddev 0.26; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,16,512,16384,512,1140.91,128,28.58,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 19.33; tg stddev 1.47; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,0,512,924.03,128,27.67,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 9.50; tg stddev 0.15; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,0,512,702.47,128,21.85,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 32.72; tg stddev 0.19; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,4096,512,903.69,128,26.53,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 22.48; tg stddev 0.26; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,16,512,16384,512,783.71,128,25.07,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 20.26; tg stddev 0.88; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,0,512,4912.05,128,111.91,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 240.22; tg stddev 0.04",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,4096,512,4254.17,128,100.56,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 113.09; tg stddev 0.10",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,16,512,16384,512,2916.49,128,77.08,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 52.09; tg stddev 0.09",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,0,512,5040.89,128,72.39,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 263.65; tg stddev 0.08",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,4096,512,4357.8,128,67.5,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 118.75; tg stddev 0.08",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,16,512,16384,512,2971.17,128,56.26,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 52.86; tg stddev 0.06",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,0,512,7449.32,128,154.54,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 601.47; tg stddev 0.08",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,4096,512,7120.95,128,147.42,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 262.51; tg stddev 0.12",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,16,512,16384,512,6020.51,128,135.25,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 228.87; tg stddev 0.13",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,0,512,7723.96,128,115.68,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 675.25; tg stddev 0.27",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,4096,512,7406.24,128,111.81,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 289.35; tg stddev 0.26",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,16,512,16384,512,6236.16,128,104.96,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 236.33; tg stddev 0.24",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,0,512,4652.29,128,101.31,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 256.98; tg stddev 0.15",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,4096,512,4541.33,128,98.29,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 133.49; tg stddev 0.13",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,16,512,16384,512,4073.31,128,92.76,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 106.97; tg stddev 0.10",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,0,512,4757.37,128,67.58,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 263.48; tg stddev 0.10",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,4096,512,4628.9,128,66.23,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 129.85; tg stddev 0.10",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,16,512,16384,512,4148.84,128,63.74,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 112.99; tg stddev 0.08",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,0,512,706.45,128,26.53,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 36.48; tg stddev 1.56; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,4096,512,650.11,128,28.33,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 13.38; tg stddev 3.69; HIGH VARIANCE in tg (>10% of mean); MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,16,512,16384,512,635.06,128,25.26,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 15.34; tg stddev 0.16; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,2548.26,128,48.4,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 112.01; tg stddev 2.75; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,1675.3,128,28.72,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 87.95; tg stddev 1.10; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,0,512,1328.67,128,22.45,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 52.89; tg stddev 0.02; MoE expert offload: --n-cpu-moe 18 (expert tensors of the first 18 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,4096,512,2468.39,128,44.11,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 44.95; tg stddev 0.16; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070,NVIDIA GeForce RTX 5070,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,16,512,16384,512,2294.97,128,43.94,3e4dfa8f3,11658,3,2026-10-16,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper PRO 3955WX 16-Cores, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 34.38; tg stddev 2.22; MoE expert offload: --n-cpu-moe 6 (expert tensors of the first 6 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,28,512,0,512,4164.98,128,89.15,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 311.50; tg stddev 0.31",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,28,512,4096,512,3737.96,128,84.14,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 230.09; tg stddev 0.32",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,28,512,16384,512,3160.99,128,83.22,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 153.91; tg stddev 0.36",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,28,512,0,512,4333.81,128,58.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 325.92; tg stddev 0.18",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,28,512,4096,512,3846.63,128,56.2,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 202.49; tg stddev 0.18",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,28,512,16384,512,3222.2,128,55.73,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 179.80; tg stddev 0.16",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,28,512,0,512,859.2,128,80.89,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 29.51; tg stddev 2.58; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,28,512,0,512,527.13,128,59.38,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 22.48; tg stddev 0.79; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,28,512,0,512,390.63,128,45.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 9.39; tg stddev 2.72; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,28,512,4096,512,859.21,128,82.06,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 10.34; tg stddev 1.37; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,28,512,16384,512,841.41,128,76.38,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 14.79; tg stddev 0.32; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,28,512,0,512,2306.74,128,52.85,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 46.30; tg stddev 0.07",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,28,512,4096,512,2199.84,128,50.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 67.90; tg stddev 0.06",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,28,512,0,512,691.0,128,78.15,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 15.82; tg stddev 7.76; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,28,512,0,512,409.73,128,55.64,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 6.48; tg stddev 1.35; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,28,512,0,512,297.07,128,39.63,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 6.94; tg stddev 1.96; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,28,512,4096,512,703.7,128,81.18,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 12.96; tg stddev 1.82; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,28,512,16384,512,643.5,128,73.25,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 5.59; tg stddev 0.60; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,28,512,0,512,5879.0,128,139.98,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 498.91; tg stddev 0.55",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,28,512,4096,512,5126.77,128,124.94,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 265.73; tg stddev 0.51",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,28,512,16384,512,3590.3,128,96.19,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 135.58; tg stddev 0.30",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,28,512,0,512,6117.64,128,93.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 548.31; tg stddev 0.36",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,28,512,4096,512,5293.45,128,86.33,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 265.62; tg stddev 0.26",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,28,512,16384,512,3697.07,128,72.01,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 140.03; tg stddev 0.18",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,28,512,0,512,8889.2,128,187.62,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1563.55; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.56",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,28,512,4096,512,8265.26,128,179.75,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 953.38; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.32",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,28,512,16384,512,6591.77,128,166.78,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 627.88; tg stddev 0.99",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,28,512,0,512,9400.32,128,143.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 1426.02; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.13",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,28,512,4096,512,8767.1,128,139.14,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 834.21; tg stddev 0.88",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,28,512,16384,512,6870.43,128,131.4,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 639.25; tg stddev 0.63",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,28,512,0,512,5871.6,128,126.03,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 520.43; tg stddev 0.80",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,28,512,4096,512,5544.27,128,122.67,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 330.41; tg stddev 0.58",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,28,512,16384,512,4729.47,128,116.36,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 290.42; tg stddev 0.50",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,28,512,0,512,6091.48,128,86.39,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 541.42; tg stddev 0.24",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,28,512,4096,512,5700.46,128,84.85,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 453.57; tg stddev 0.28",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,28,512,16384,512,4837.67,128,81.83,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 299.01; tg stddev 0.21",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,28,512,0,512,405.88,128,68.35,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 15.06; tg stddev 3.14; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,28,512,0,512,289.72,128,53.05,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 10.11; tg stddev 1.07; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,28,512,4096,512,391.83,128,69.51,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 3.60; tg stddev 0.83; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,28,512,16384,512,381.88,128,53.73,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 1.78; tg stddev 3.54; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,28,512,0,512,9595.62,128,236.89,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1354.11; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.18",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,28,512,4096,512,8476.64,128,217.64,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1132.88; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.58",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,28,512,16384,512,6736.99,128,208.7,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 435.92; tg stddev 3.52",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,28,512,0,512,9839.06,128,223.82,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 1503.87; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.47",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,28,512,4096,512,8778.77,128,207.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 821.31; tg stddev 1.59",llama.cpp | |
| rtx-5070ti,NVIDIA GeForce RTX 5070 Ti,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,28,512,16384,512,6858.55,128,198.93,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 519.00; tg stddev 2.83",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,0,512,4597.77,128,94.8,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 405.33; tg stddev 0.31",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,4096,512,4194.13,128,89.89,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 185.59; tg stddev 0.25",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,16384,512,3586.7,128,88.53,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 126.99; tg stddev 0.30",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,0,512,4712.82,128,61.81,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 431.11; tg stddev 0.13",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,4096,512,4312.0,128,59.65,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 188.26; tg stddev 0.11",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,16384,512,3673.12,128,59.08,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 135.15; tg stddev 0.11",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,0,512,1304.53,128,65.8,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 89.70; tg stddev 0.87; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,0,512,797.81,128,43.84,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 37.59; tg stddev 0.15; MoE expert offload: --n-cpu-moe 15 (expert tensors of the first 15 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,0,512,647.14,128,32.65,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 31.08; tg stddev 0.34; MoE expert offload: --n-cpu-moe 22 (expert tensors of the first 22 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,4096,512,1374.06,128,64.04,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 19.08; tg stddev 0.81; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,16384,512,1290.96,128,61.91,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 106.29; tg stddev 0.93; MoE expert offload: --n-cpu-moe 8 (expert tensors of the first 8 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,0,512,2581.12,128,58.64,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 96.69; tg stddev 0.05",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,4096,512,2425.44,128,56.04,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 75.73; tg stddev 0.05",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,0,512,1086.09,128,66.93,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 111.01; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.20; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,0,512,678.86,128,39.46,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 23.06; tg stddev 0.50; MoE expert offload: --n-cpu-moe 24 (expert tensors of the first 24 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,0,512,492.88,128,28.43,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 22.40; tg stddev 0.10; MoE expert offload: --n-cpu-moe 36 (expert tensors of the first 36 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,4096,512,1108.53,128,63.29,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 41.78; tg stddev 0.91; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,16384,512,979.49,128,58.52,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 54.77; tg stddev 0.98; MoE expert offload: --n-cpu-moe 12 (expert tensors of the first 12 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,0,512,6837.85,128,149.38,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 655.67; tg stddev 0.51",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,4096,512,6014.0,128,134.92,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 298.07; tg stddev 0.43",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,16384,512,4034.12,128,105.49,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 130.59; tg stddev 0.28",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,0,512,7048.14,128,98.89,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 806.60; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.20",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,4096,512,6219.92,128,92.42,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 294.50; tg stddev 0.19",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,16384,512,4143.61,128,77.63,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 128.86; tg stddev 0.14",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,0,512,10057.3,128,200.46,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1520.80; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.12",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,4096,512,9750.07,128,191.66,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 650.13; tg stddev 0.93",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,16384,512,8287.75,128,178.74,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 551.23; tg stddev 0.93",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,0,512,10589.57,128,153.35,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 1738.31; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.76",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,4096,512,10214.8,128,148.04,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 712.88; tg stddev 0.56",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,16384,512,8646.12,128,140.68,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 604.92; tg stddev 0.58",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,0,512,6543.41,128,136.14,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 711.02; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.58",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,4096,512,6403.26,128,131.86,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 342.31; tg stddev 0.50",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,16384,512,5739.87,128,125.85,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 292.76; tg stddev 0.43",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,0,512,6790.2,128,92.26,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 750.41; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.25",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,4096,512,6694.94,128,90.33,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 309.31; tg stddev 0.21",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,16384,512,5903.9,128,87.47,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 365.58; tg stddev 0.20",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,0,512,680.53,128,56.1,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 53.20; tg stddev 0.49; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,0,512,501.12,128,41.91,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 4.30; tg stddev 0.44; MoE expert offload: --n-cpu-moe 30 (expert tensors of the first 30 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,4096,512,688.57,128,56.48,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 44.20; tg stddev 0.39; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,16384,512,647.41,128,52.81,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 27.30; tg stddev 0.48; MoE expert offload: --n-cpu-moe 20 (expert tensors of the first 20 layers in system RAM); excluded from site medians",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,0,512,10825.61,128,253.49,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1577.95; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.35",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,4096,512,10031.52,128,234.1,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 726.36; tg stddev 1.11",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,16384,512,8156.33,128,221.77,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 479.80; tg stddev 4.71",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,0,512,11105.78,128,238.83,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 1646.97; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.12",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,4096,512,10297.11,128,221.8,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 754.49; tg stddev 1.20",llama.cpp | |
| rtx-5080,NVIDIA GeForce RTX 5080,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,16384,512,8196.62,128,211.79,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD EPYC 7D12 32-Core Processor, RAM 125 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 503.50; tg stddev 2.65",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,0,512,8523.97,128,152.43,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 1285.03; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.68",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,4096,512,7893.63,128,145.24,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 611.05; tg stddev 0.59",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,16384,512,6818.48,128,143.16,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 432.67; tg stddev 0.59",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,0,512,8805.63,128,103.67,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 1377.66; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.38",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,4096,512,8095.07,128,100.14,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 655.07; tg stddev 0.30",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,16384,512,6921.91,128,99.28,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 455.79; tg stddev 0.28",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,0,512,11431.56,128,231.41,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 2287.16; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.95",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,4096,512,10288.68,128,216.39,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 1329.66; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.02",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,16384,512,9261.99,128,213.02,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 851.28; tg stddev 1.95",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,0,512,11322.19,128,217.48,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 2357.32; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.82",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,4096,512,10335.09,128,204.47,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 1170.46; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.64",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,16384,512,9257.66,128,201.66,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 780.33; tg stddev 1.60",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,0,512,3841.2,128,71.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 293.07; tg stddev 0.23",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,4096,512,3606.52,128,67.18,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 150.53; tg stddev 0.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,16384,512,2998.41,128,64.61,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 102.48; tg stddev 0.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,0,512,5184.84,128,101.14,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 340.21; tg stddev 0.24",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,4096,512,5053.08,128,96.2,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 207.05; tg stddev 0.20",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,16384,512,4239.39,128,85.81,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 156.99; tg stddev 0.17",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,0,512,5246.27,128,63.85,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 294.24; tg stddev 0.10",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,4096,512,5137.43,128,61.85,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 191.25; tg stddev 0.08",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,16384,512,4306.26,128,57.39,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 138.60; tg stddev 0.07",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,0,512,10437.02,128,346.27,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 2091.65; HIGH VARIANCE in pp (>10% of mean); tg stddev 3.49",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,4096,512,9270.48,128,287.5,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 1121.58; HIGH VARIANCE in pp (>10% of mean); tg stddev 3.16",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,16384,512,6667.26,128,245.24,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 579.52; tg stddev 3.53",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,0,512,12303.0,128,242.38,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 2128.17; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.15",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,4096,512,11282.87,128,217.4,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 1007.67; tg stddev 0.88",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,16384,512,8220.86,128,173.77,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 553.60; tg stddev 0.77",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,0,512,12638.69,128,166.54,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 2183.60; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.60",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,4096,512,11507.49,128,154.58,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 999.70; tg stddev 0.54",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,16384,512,8350.77,128,131.31,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 570.47; tg stddev 0.48",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,0,512,16035.6,128,313.64,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 3670.77; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.63",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,4096,512,15709.28,128,299.6,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1921.56; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.80",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,16384,512,13499.68,128,282.36,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1713.22; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.37",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,0,512,17193.33,128,244.56,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 4232.67; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.10",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,4096,512,16763.77,128,236.54,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 2182.34; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.59",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,16384,512,14262.66,128,226.18,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 1800.63; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.44",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,0,512,11720.74,128,220.46,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 1946.30; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,4096,512,11596.47,128,213.8,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 1070.36; tg stddev 1.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,16384,512,10365.19,128,205.18,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 909.54; tg stddev 1.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,0,512,11960.03,128,156.01,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 2145.91; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.80",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,4096,512,11880.68,128,152.13,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 1095.51; tg stddev 0.64",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,16384,512,10618.36,128,147.79,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 915.85; tg stddev 0.60",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,0,512,4084.61,128,77.61,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 487.41; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.38",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,4096,512,4065.95,128,75.78,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 265.79; tg stddev 0.30",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,16384,512,3623.67,128,74.13,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 220.20; tg stddev 0.31",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,0,512,4064.01,128,52.36,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 505.99; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.22",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,4096,512,4031.04,128,51.44,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 286.82; tg stddev 0.15",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,16384,512,3591.7,128,50.7,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 234.64; tg stddev 0.16",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,0,512,9602.39,128,256.87,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 2344.81; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.96",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,4096,512,9249.0,128,243.77,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 1266.53; HIGH VARIANCE in pp (>10% of mean); tg stddev 3.17",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,16384,512,8408.35,128,247.21,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 961.53; HIGH VARIANCE in pp (>10% of mean); tg stddev 3.50",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,0,512,17630.25,128,393.3,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 3787.81; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.62",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,4096,512,16548.12,128,354.98,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1978.06; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.80",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,16384,512,13689.98,128,345.21,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1460.81; HIGH VARIANCE in pp (>10% of mean); tg stddev 6.75",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,0,512,18400.2,128,374.72,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 4059.14; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.73",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,4096,512,17289.11,128,339.33,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 2009.92; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.30",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,16384,512,14069.65,128,331.04,3e4dfa8f3,11658,3,2026-10-17,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.84, vast.ai, CPU AMD EPYC 7532 32-Core Processor, RAM 188 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 1415.31; HIGH VARIANCE in pp (>10% of mean); tg stddev 5.96",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,24,512,0,512,4629.32,128,82.51,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 206.30; tg stddev 0.23",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,24,512,4096,512,4502.01,128,81.57,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 201.46; tg stddev 0.23",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,24,512,16384,512,4031.47,128,78.27,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 153.92; tg stddev 0.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,24,512,32768,512,3316.89,128,74.5,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 112.82; tg stddev 0.18",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q4_K_M,27.78,15.32,1,99,24,512,65536,512,2107.34,128,67.89,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_M.gguf; pp stddev 32.23; tg stddev 0.17",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q8_0,27.78,27.04,1,99,24,512,0,512,4378.16,128,52.98,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q8_0.gguf; pp stddev 336.02; tg stddev 0.09",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q8_0,27.78,27.04,1,99,24,512,4096,512,4293.76,128,52.53,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q8_0.gguf; pp stddev 198.76; tg stddev 0.19",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q8_0,27.78,27.04,1,99,24,512,16384,512,3886.89,128,51.28,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q8_0.gguf; pp stddev 152.00; tg stddev 0.10",llama.cpp | |
| rtx-5090,NVIDIA GeForce RTX 5090,1,CUDA,Qwen3.8 27B,Q8_0,27.78,27.04,1,99,24,512,32768,512,3232.42,128,49.6,3e4dfa8f3,11658,3,2026-10-26,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.99.02, vast.ai, CPU AMD Ryzen Threadripper 9960X 24-Cores, RAM 251 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q8_0.gguf; pp stddev 103.57; tg stddev 0.09",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,GLM-4.5-Air,Q4_K_M,110.47,67.96,1,99,32,512,0,512,3634.78,128,138.28,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'glm4moe 106B.A12B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/GLM-4.5-Air-GGUF/Q4_K_M/GLM-4.5-Air-Q4_K_M-00001-of-00002.gguf; pp stddev 162.61; tg stddev 0.37",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,GLM-4.5-Air,Q4_K_M,110.47,67.96,1,99,32,512,4096,512,3304.59,128,117.91,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'glm4moe 106B.A12B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/GLM-4.5-Air-GGUF/Q4_K_M/GLM-4.5-Air-Q4_K_M-00001-of-00002.gguf; pp stddev 114.89; tg stddev 0.39",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,GLM-4.5-Air,Q4_K_M,110.47,67.96,1,99,32,512,16384,512,2264.58,128,104.48,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'glm4moe 106B.A12B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/GLM-4.5-Air-GGUF/Q4_K_M/GLM-4.5-Air-Q4_K_M-00001-of-00002.gguf; pp stddev 51.11; tg stddev 0.42",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,0,512,9191.99,128,150.41,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 1128.27; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.61",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,4096,512,8503.89,128,143.62,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 633.72; tg stddev 0.53",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q4_K_M,11.96,6.62,1,99,32,512,16384,512,7495.32,128,141.65,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q4_K_M.gguf; pp stddev 430.16; tg stddev 0.51",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,0,512,9336.32,128,101.19,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 1347.70; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.24",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,4096,512,8675.29,128,97.99,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 660.61; tg stddev 0.25",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 12B,Q8_0,11.96,11.78,1,99,32,512,16384,512,7628.86,128,97.04,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 ?B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-Q8_0.gguf; pp stddev 482.73; tg stddev 0.26",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,0,512,12584.24,128,229.63,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 2224.04; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.56",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,4096,512,11476.89,128,214.44,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 1043.11; tg stddev 1.40",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q4_K_M,25.81,15.77,1,99,32,512,16384,512,10349.35,128,211.37,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; pp stddev 867.59; tg stddev 1.32",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,0,512,12320.37,128,213.84,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 2160.40; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.35",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,4096,512,11300.81,128,200.31,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 1108.28; tg stddev 1.27",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,32,512,16384,512,10293.33,128,197.91,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 752.15; tg stddev 1.07",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,0,512,3910.45,128,69.38,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 133.54; tg stddev 0.12",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,4096,512,3773.84,128,65.53,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 155.42; tg stddev 0.11",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q4_K_M,31.27,17.05,1,99,32,512,16384,512,3162.32,128,63.41,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q4_K_M.gguf; pp stddev 98.66; tg stddev 0.09",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,32,512,0,512,3889.02,128,43.91,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 125.73; tg stddev 0.05",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,32,512,4096,512,3752.68,128,42.26,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 136.01; tg stddev 0.05",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,32,512,16384,512,3134.04,128,41.37,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 99.23; tg stddev 0.06",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,0,512,5393.81,128,97.84,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 161.68; tg stddev 0.11",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,4096,512,5312.16,128,93.28,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 184.31; tg stddev 0.16",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q4_K_M,24.01,13.34,1,99,32,512,16384,512,4522.27,128,83.34,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf; pp stddev 147.84; tg stddev 0.08",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,0,512,5547.25,128,62.15,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 149.67; tg stddev 0.06",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,4096,512,5492.47,128,60.32,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 172.68; tg stddev 0.06",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,32,512,16384,512,4678.82,128,56.06,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 122.71; tg stddev 0.06",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,0,512,11393.13,128,346.25,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 1938.35; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.56",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,4096,512,10436.91,128,293.11,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 839.44; tg stddev 2.34",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q4_K_M,30.53,17.28,1,99,32,512,16384,512,7645.71,128,246.12,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf; pp stddev 509.80; tg stddev 2.14",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,32,512,0,512,11602.57,128,275.5,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 1954.77; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.37",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,32,512,4096,512,10509.24,128,239.79,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 873.42; tg stddev 1.99",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,32,512,16384,512,7672.51,128,207.65,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 571.90; tg stddev 1.61",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,0,512,13811.9,128,239.61,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 1956.75; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.80",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,4096,512,12279.31,128,215.77,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 922.47; tg stddev 0.72",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q4_K_M,8.77,4.68,1,99,32,512,16384,512,9216.91,128,172.41,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q4_K_M.gguf; pp stddev 461.21; tg stddev 0.48",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,0,512,14173.59,128,163.23,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 2138.70; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.41",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,4096,512,12674.81,128,152.19,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 941.53; tg stddev 0.37",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3-VL 8B,Q8_0,8.77,8.11,1,99,32,512,16384,512,9331.35,128,129.7,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen3vl 8B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3-VL-8B-Instruct-GGUF/Qwen3-VL-8B-Instruct-Q8_0.gguf; pp stddev 572.56; tg stddev 0.31",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,0,512,17977.45,128,314.88,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 3914.01; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.01",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,4096,512,17462.41,128,299.13,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1728.78; tg stddev 1.84",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q4_K_M,4.66,2.54,1,99,32,512,16384,512,15185.03,128,283.39,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q4_K_M.gguf; pp stddev 1408.44; tg stddev 1.71",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,0,512,19065.9,128,243.63,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 4108.23; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.23",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,4096,512,18319.75,128,233.82,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 2340.48; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.27",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 4B,Q8_0,4.66,4.16,1,99,32,512,16384,512,15879.53,128,225.05,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 4B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf; pp stddev 1588.95; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.07",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,0,512,13098.49,128,219.59,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 1970.36; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.11",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,4096,512,12837.03,128,211.48,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 1135.08; tg stddev 0.93",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q4_K_M,9.65,5.28,1,99,32,512,16384,512,11578.67,128,204.07,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.gguf; pp stddev 877.14; tg stddev 0.95",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,0,512,13411.39,128,153.43,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 2156.15; HIGH VARIANCE in pp (>10% of mean); tg stddev 0.52",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,4096,512,13046.27,128,149.39,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 1212.20; tg stddev 0.47",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.5 9B,Q8_0,9.65,8.86,1,99,32,512,16384,512,11778.45,128,145.84,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 9B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf; pp stddev 916.79; tg stddev 0.48",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,0,512,4353.24,128,76.11,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 339.48; tg stddev 0.27",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,4096,512,4432.81,128,74.02,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 242.48; tg stddev 0.20",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q4_K_M,27.78,15.65,1,99,32,512,16384,512,3965.62,128,72.45,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf; pp stddev 193.27; tg stddev 0.18",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,0,512,4416.94,128,51.1,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 420.23; tg stddev 0.13",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,4096,512,4496.24,128,50.08,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 260.64; tg stddev 0.16",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,32,512,16384,512,4022.41,128,49.44,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 219.64; tg stddev 0.10",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,0,512,10646.5,128,260.07,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 2175.74; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.55",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,4096,512,10237.43,128,242.63,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 1216.38; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.99",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q4_K_M,35.95,20.6,1,99,32,512,16384,512,9363.49,128,245.59,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; pp stddev 920.28; tg stddev 2.09",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,32,512,0,512,10165.49,128,246.84,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 2067.63; HIGH VARIANCE in pp (>10% of mean); tg stddev 3.27",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,32,512,4096,512,9802.7,128,233.1,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 1139.17; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.83",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,32,512,16384,512,9013.7,128,235.19,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 864.29; tg stddev 1.81",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q4_K_M,116.83,58.45,1,99,32,512,0,512,9016.89,128,276.94,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q4_K_M/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf; pp stddev 1314.02; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.65",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q4_K_M,116.83,58.45,1,99,32,512,4096,512,8761.2,128,249.85,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q4_K_M/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf; pp stddev 730.86; tg stddev 1.47",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q4_K_M,116.83,58.45,1,99,32,512,16384,512,7707.92,128,242.46,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q4_K_M/gpt-oss-120b-Q4_K_M-00001-of-00002.gguf; pp stddev 533.66; tg stddev 4.31",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q8_0,116.83,59.02,1,99,32,512,0,512,9341.7,128,260.76,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q8_0/gpt-oss-120b-Q8_0-00001-of-00002.gguf; pp stddev 1273.44; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.53",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q8_0,116.83,59.02,1,99,32,512,4096,512,8968.1,128,237.07,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q8_0/gpt-oss-120b-Q8_0-00001-of-00002.gguf; pp stddev 725.82; tg stddev 1.45",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 120B,Q8_0,116.83,59.02,1,99,32,512,16384,512,7870.16,128,231.41,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 120B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-120b-GGUF/Q8_0/gpt-oss-120b-Q8_0-00001-of-00002.gguf; pp stddev 500.23; tg stddev 2.67",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,0,512,19560.11,128,395.98,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 3559.13; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.15",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,4096,512,18372.03,128,359.91,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1929.49; HIGH VARIANCE in pp (>10% of mean); tg stddev 2.21",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q4_K_M,20.91,10.81,1,99,32,512,16384,512,15172.33,128,349.43,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q4_K - Medium'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf; pp stddev 1303.63; tg stddev 5.76",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,0,512,20480.04,128,375.71,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 4026.30; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.97",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,4096,512,19177.05,128,342.79,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 2106.76; HIGH VARIANCE in pp (>10% of mean); tg stddev 1.70",llama.cpp | |
| rtx-pro-6000,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,1,CUDA,gpt-oss 20B,Q8_0,20.91,11.27,1,99,32,512,16384,512,15721.88,128,332.3,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.91.07, vast.ai, CPU AMD EPYC 7543 32-Core Processor, RAM 503 GiB; llama-bench model_type 'gpt-oss 20B Q8_0'; -r 3; GPU 0 only; GGUF unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.gguf; pp stddev 1230.16; tg stddev 4.67",llama.cpp | |