bottlecapai/ThinkingCap-gemma-4-26B-A4B-it
Image-Text-to-Text • 26B • Updated • 804 • 26
ThinkingCap gemma-4-26B-A4B-it: bf16 for vLLM and Transformers, GGUF (MXFP4_MOE recommended) for llama.cpp.
Note 🎓 bf16 reference, vLLM/Transformers, 51.6 GB; served on one GB300 in our evals. Against the base gemma-4-26B-A4B-it: 38.9% fewer reasoning tokens on average (14% to 69% by benchmark) at 75.0% average accuracy vs 75.2%. The baseline for the GGUF below.
Note 🦙 GGUF with vision projector, llama.cpp/LM Studio/Ollama: MXFP4_MOE 16.6 GB (recommended), IQ4_XS 13.6, Q4_K_M 17.0, Q6_K 23.3, Q8_0 26.9; attention, dense block and embeddings stay Q8_0 in every file. MXFP4_MOE evals vs bf16 on nine benchmarks: accuracy −3.3 to +10.7 pp, the two significant changes both gains (HMMT Feb 2026, RealWorldQA); mean thinking −25% to +17%.