MLX Model Conversion

Convert Mistral Devstral model to MLX format for Apple Silicon.

Makefile Targets

install

Install project dependencies using uv sync

convert

Run the model conversion script (convert_model.py) to convert the Hugging Face model to MLX format with quantization

run

Execute the converted MLX model (run_model.py)

clean

Remove generated directories and cache files:

• Devstral-Small-2-24B-Instruct-2512-mlx
• .venv
• __pycache__
• .ruff_cache

convert_model.py

The convert_model.py script converts the Mistral Devstral model from Hugging Face to MLX format optimized for Apple Silicon. It uses the mlx_lm library to convert the model with quantization enabled, reducing model size while maintaining performance.

Model details:

• Source: mistralai/Devstral-Small-2-24B-Instruct-2512
• Output: ./Devstral-Small-2-24B-Instruct-2512-mlx
• Upload target: mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit