Modelpublic

deresolution/Qwen3.8-27B-DFlash2-mxfp4

sourceHugging Faceapache-2.0updated 28d agoView on Hugging Face

1likes1.9kdownloads

Model Card

Qwen3.8-27B DFlash2 MXFP4 MLX

An unofficial MLX MXFP4 conversion of `incoai/Qwen3.8-27B-DFlash2`, produced for Apple Silicon speculative decoding with oMLX. It is a quantized conversion, not a fine-tune, merge, retraining, or new DFlash model. It is not affiliated with or endorsed by Inco AI or Qwen.

This is not a standalone chat model. It is a DFlash2 draft sidecar for the matching `Qwen/Qwen3.8-27B` target architecture. The large target model remains responsible for verifying every drafted token and producing the final output.

Conversion

Setting	Value
Source	`incoai/Qwen3.8-27B-DFlash2`
Conversion tool	`mlx_vlm.convert`
Weight format	MXFP4
Nominal bits	4
Group size	32
Non-quantized dtype	BF16
Resulting disk size	approximately 1 GB

The conversion command was:

bash

mlx_vlm.convert \
  --hf-path incoai/Qwen3.8-27B-DFlash2 \
  --quantize \
  --q-bits 4 \
  --q-mode mxfp4 \
  --q-group-size 32 \
  --dtype bfloat16 \
  --mlx-path Qwen3.8-27B-DFlash2-mxfp4

What DFlash2 does

DFlash2 is a speculative-decoding drafter. It predicts a short block of likely next tokens; the full Qwen3.8-27B target verifies them and accepts only the tokens that match. This can reduce the number of expensive target-model decode passes without changing the target model's authority over the output.

Actual speedup depends on the prompt, sampling settings, context length, and draft acceptance rate. It does not improve prompt-prefill speed.

Use with oMLX

Use this as the DFlash draft for a separately downloaded Qwen3.8-27B target, such as `deresolution/Qwen3.8-27B-mxfp4`.

In the target model's oMLX profile:

1.Enable DFlash.
2.Choose this model as the draft model.
3.Disable runtime draft quantization because this checkpoint is already MXFP4-quantized.
4.Set DFlash block size to 5 for this quantized target/draft pairing.
5.Do not enable Lightning MTP and DFlash at the same time.

Keep this sidecar unloaded as a standalone model in oMLX's model library. The target's DFlash engine attaches it internally when the target is loaded.

Compatibility

This conversion was tested as the draft sidecar for the Qwen3.8-27B MXFP4 target above on a 32 GB Apple Silicon Mac using oMLX. It is not a general purpose MLX chat checkpoint and is not expected to work as a draft for other Qwen families or unrelated targets.

For the DFlash algorithm and upstream benchmark details, see the DFlash repository, the upstream DFlash2 model card, and oMLX's experimental DFlash integration guide.

License and attribution

This derivative conversion retains the upstream Apache License 2.0. Please refer to the upstream DFlash2 model card for intended use, limitations, evaluation, and citation information.