unigilby/MiniMax-M3-MLX-MXFP4-imatrix

MiniMax-M3 — MLX MXFP4 (imatrix-calibrated)

Mixed-precision 4-bit MLX quantization of MiniMaxAI/MiniMax-M3, calibrated with an importance matrix (imatrix). ~217 GB on disk.

Quantization map

Scope	Mode	Bits	Group size
Body default	affine	4	64
116 layers	mxfp4 (microscaling FP4)	4	32
365 sensitive layers (embed_tokens, lm_head, select MLP projections)	mxfp8	8	32

Calibration: imatrix-guided quantization (oMLX oQe pipeline). In our internal evaluation the imatrix variant recovered ~2.3% perplexity versus the non-calibrated 4-bit quant of the same model.

Usage (MLX / Apple Silicon)

Requires a recent MLX + mlx-lm with MiniMax-M3 (MiniMaxM3SparseForConditionalGeneration) support. Runs on Apple Silicon with ≥256 GB unified memory (tested on a 512 GB Mac Studio M3 Ultra).

from mlx_lm import load, generate
model, tokenizer = load("unigilby/MiniMax-M3-MLX-MXFP4-imatrix")
print(generate(model, tokenizer, prompt="Hello", max_tokens=64))

License

Inherits the MiniMax model license from the base repository — see the license link above.