Qwen3.5-27B-DFlash

Paper | GitHub | Blog

This model is still under training.

DFlash is a novel speculative decoding method that utilizes a lightweight block diffusion model for drafting. It enables efficient, high-quality parallel drafting that pushes the limits of inference speed.

This model is the drafter component. It must be used in conjunction with the target model Qwen/Qwen3.5-27B. It was trained with a context length of 4096 tokens.

🚀 Quick Start

SGLang

Installation

uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/16818/head#subdirectory=python"

Inference

python -m sglang.launch_server \
    --model-path Qwen/Qwen3.5-27B \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3.5-27B-DFlash \
    --speculative-num-draft-tokens 16 \
    --tp-size 1 \
    --attention-backend fa3 \
    --mem-fraction-static 0.75 \
    --mamba-scheduler-strategy extra_buffer \
    --trust-remote-code

Note: For long-context or agentic usage, consider adding --speculative-dflash-draft-window-size WINDOW_SIZE to enable sliding-window attention for the draft model.

vLLM

Thanks to the community and all contributors! Check out the following PRs to see how to run DFlash on vLLM: #36847 and #36767.

Early Results

• Thinking: enabled
• Max new tokens: 4096
• Block size: 16
• 0.5 Epoch

  Dataset	Accept Length
  GSM8K	5.92
  Math500	6.49
  HumanEval	7.26
  MBPP	5.75
  MT-Bench	4.47
  Alpaca	4.15