Qwen3-Coder-Next-DFlash

Paper | GitHub | Blog

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-Coder-Next or its FP8 variant. 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-Coder-Next \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3-Coder-Next-DFlash \
    --speculative-num-draft-tokens 16 \
    --tp-size 1 \
    --dtype bfloat16 \
    --attention-backend fa3 \
    --mem-fraction-static 0.75 \
    --trust-remote-code \
    --mamba-scheduler-strategy extra_buffer \
    --tool-call-parser qwen3_coder

Note: For long-context or agentic usage (such as OpenClaw or Claude Code), consider adding --speculative-dflash-draft-window-size WINDOW_SIZE to enable sliding-window attention for the draft model. Because the draft model is only trained on 4K context, this often improves performance on very long context (50K+ tokens).

Results

• Max new tokens: 4096
• Block size: 16

  Dataset	Accept Length
  HumanEval	xxx
  MBPP	xxx
  LiveCodeBench	xxx