bouroo/Qwopus3.8-27B-Coder-Flash-MTP-mxfp4

Qwopus3.8-27B-Coder-Flash MTP MXFP4

MXFP4 (4-bit, group 32) MLX conversion of the MTP drafter of Jackrong/Qwopus3.8-27B-Coder-Flash.

This is not a standalone model. It holds only the Multi-Token Prediction head (mtp.*) of the fine-tune, quantized to MXFP4, and is used as the draft model alongside the matching target, bouroo/Qwopus3.8-27B-Coder-Flash-mxfp4. The target supplies the token embeddings and the language-model head.

Use with mlx-vlm

pip install -U mlx-vlm

mlx_vlm.generate \
  --model bouroo/Qwopus3.8-27B-Coder-Flash-mxfp4 \
  --draft-model bouroo/Qwopus3.8-27B-Coder-Flash-MTP-mxfp4 \
  --prompt "Write a Python function that merges overlapping intervals. Include tests." \
  --max-tokens 512 \
  --enable-thinking

--draft-kind is auto-detected from model_type: qwen3_5_mtp. The default block size is 3. Measured on an M4 Pro: 2.78 tokens accepted per round, 89.5% of the drafted tokens accepted.

Files

File	What
model.safetensors	The MTP drafter: fc, one decoder layer, the shared norms (235 MB)
config.json	model_type: qwen3_5_mtp, block_size: 3, the target's text_config, MXFP4 quantization
tokenizer.json, tokenizer_config.json	The target's tokenizer, copied verbatim

Quantization

The drafter's 2-D weights (the fc projection, the layer's attention and MLP projections) are MXFP4 (E2M1 codes, group size 32, one uint8 power-of-two scale per group). Its norm weights are stored as 1 + w, matching transformers' Qwen3_5RMSNorm, and the loader's own sanitize pass leaves them unchanged.

The layout mirrors mlx-community/Qwen3.8-27B-MTP-mxfp4, the MXFP4 drafter of the base Qwen3.8-27B.

License and limitations

Apache-2.0, inherited from the upstream checkpoint. The upstream model card's stated limitations apply.