Modelpublic

aws-prototyping/Qwen3-Coder-30B-A3B-Instruct_MXFP4

sourceHugging Faceapache-2.0updated 9mo agoView on Hugging Face

0likes41downloads

Model Card

Qwen3-Coder-30B-A3B-Instruct_MXFP4

This checkpoint is a variant of Qwen3-Coder-30B-A3B-Instruct, where expert weights have been quantized to MXFP4 format similarly to gpt-oss-20b and gpt-oss-120b.

For quantizing weights we used the function downcast_to_mxfp from triton-kernels.

The checkpoint might come with a small drop in accuracy, but has ~68% size reduction compared to the original BF16 checkpoint.

Model	GSM8K (strict-match)	GSM8K (flexible-extract)
Qwen3-Coder-30B-A3B-Instruct (BF16)	90.67% ± 0.80%	89.92% ± 0.83%
Qwen3-Coder-30B-A3B-Instruct_MXFP4	89.76% ± 0.83%	88.70% ± 0.87%

Model	Size	Reduction
Qwen3-Coder-30B-A3B-Instruct (BF16)	57 GB	-
Qwen3-Coder-30B-A3B-Instruct_MXFP4	18 GB	68% smaller