huggingface-lc/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit-mxfp4

Qwen3-Coder-30B-A3B-Instruct (MLX, 4-bit, mxfp4)

This model is an MLX-converted version of Qwen/Qwen3-Coder-30B-A3B-Instruct, using mlx_lm version 0.28.2, optimized for Apple Silicon. It uses 4-bit quantization with the following settings:

• Quantization: 4-bit with --q-mode mxfp4 for efficient performance on Apple Silicon.
• Group size: 32 (--q-group-size 32) for improved precision with moderate memory usage (~22-27 GB).
• Context: Supports up to 262k tokens natively, optimized for agentic coding tasks and large-scale code repository analysis.

Conversion Command

mlx_lm.convert \
    --hf-path Qwen/Qwen3-Coder-30B-A3B-Instruct \
    -q \
    --q-bits 4 \
    --q-mode mxfp4 \
    --q-group-size 32 \
    --mlx-path ./qwen3-coder-30b-mlx