Loading unsloth/Qwen3.5-4B
Merging the jica98/qwen3.5-4B-super-coder LoRA adapter with Qwen3_5ForConditionalGeneration
Converting the merged Hugging Face model to MLX
Quantizing to 4-bit with mlx-lm No additional training was done.

Use

bash

pip install -U mlx-lm
mlx_lm.generate \
  --model snsnc/qwen3.5-4B-super-coder-MLX-4bit \
  --max-tokens 2048 \
  --temp 0.7 \
  --top-p 0.8 \
  --top-k 20 \
  --chat-template-config '{"enable_thinking":false}' \
  --prompt "Write only executable Python code. Implement an async token-bucket rate limiter."

Thinking is enabled by the upstream chat template unless explicitly disabled.