Qwen3.5-9b-Sushi-Coder-RL-MLX

Lineage

• Base model lineage: bigatuna/Qwen3.5-9b-Sushi-Coder
• RL model: bigatuna/Qwen3.5-9b-Sushi-Coder-RL
• MLX export source: this repo is a 4-bit MLX conversion of the RL checkpoint above
• RL pipeline: NousResearch/atropos

Training

The upstream SFT model was trained with Unsloth on:

• nohurry/Opus-4.6-Reasoning-3000x-filtered
• open-r1/codeforces-cots

The RL stage was then run for coding with NousResearch/hermes-agent using NousResearch/atropos.

During that run, vLLM was patched with vllm-project/vllm PR #36395, fix(lora): add bounds checking for TP configurations, to address the LoRA tensor-parallel bounds issue.

Conversion

This repo contains an MLX export for Apple Silicon generated from the original Hugging Face safetensors checkpoint, not from the GGUF release.

• Format: MLX
• Quantization: 4-bit affine
• Conversion stack: mlx-vlm

Files

• model-00001-of-00002.safetensors
• model-00002-of-00002.safetensors
• model.safetensors.index.json
• config.json
• processor_config.json
• tokenizer.json
• tokenizer_config.json
• generation_config.json
• chat_template.jinja

Usage Note

This is an MLX multimodal export intended for Apple Silicon. Use it with mlx-vlm, not llama.cpp.

Quick Start

Install:

pip install -U mlx-vlm

Text generation:

mlx_vlm.generate \
  --model bigatuna/Qwen3.5-9b-Sushi-Coder-RL-MLX \
  --prompt "Write a Python function that parses a CSV file into dataclasses." \
  --max-tokens 512

Image + text:

mlx_vlm.generate \
  --model bigatuna/Qwen3.5-9b-Sushi-Coder-RL-MLX \
  --image /path/to/image.png \
  --prompt "Describe the bug in this screenshot and suggest a fix." \
  --max-tokens 512

Metadata

• License: Apache-2.0
• Architecture: Qwen 3.5
• Format: MLX
• Tags: mlx, mlx-vlm, apple-silicon, multimodal, code, rl