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

Qwopus3.8-27B-Coder-Flash MXFP4

MXFP4 (4-bit, group 32) MLX conversion of Jackrong/Qwopus3.8-27B-Coder-Flash, a coding-oriented fine-tune of Qwen3.8-27B.

The upstream checkpoint is BF16 (56 GB). This repository is the same model quantized to MXFP4 for MLX, 14.2 GB in total, with the vision tower, the Gated DeltaNet state and the norms kept at BF16.

Use it with the matching MTP drafter in bouroo/Qwopus3.8-27B-Coder-Flash-MTP-mxfp4 for speculative decoding.

Files

File	What
model-00001..00004.safetensors	MXFP4 weights (498 quantized modules) plus BF16 vision tower, norms and GDN state
model.safetensors.index.json	Tensor-to-shard map
config.json	Upstream config plus quantization: {group_size: 32, bits: 4, mode: mxfp4}
chat_template.jinja, tokenizer.json, tokenizer_config.json	Chat template and tokenizer

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

The draft model is detected as qwen3_5_mtp automatically; --draft-block-size 3 is the default. Measured on an M4 Pro: 2.78 tokens accepted per round (89.5% of the drafted tokens accepted).

Quantization

Every 2-D weight of the language model is MXFP4 (E2M1 codes, group size 32, one uint8 power-of-two scale per group). The vision tower, the RMSNorm weights, the Gated DeltaNet A_log, dt_bias, conv1d and the token table stay BF16, matching the layout of mlx-community/Qwen3.8-27B-mxfp4.

The norm weights are stored as 1 + w, the convention transformers' Qwen3_5RMSNorm uses (output * (1 + weight)), so the loader's own sanitize pass leaves them unchanged.

Provenance

Converted from Jackrong/Qwopus3.8-27B-Coder-Flash with mlx_vlm's quantization conventions, one source shard at a time, so the 56 GB BF16 checkpoint is never materialised in memory. Verified by generating with the converted weights and by loading it as an MTP draft's target.

License and limitations

Apache-2.0, inherited from the upstream checkpoint. The upstream model card's stated limitations apply: the reported evaluations are small local runs, and generated code and tool actions can be incorrect. See the upstream model card for the full description, the benchmark details and the intended use.