Qwen3.5-9B-DFlash

Paper | Github | Blog

This DFlash draft model is a joint retrain from Z-Lab and Modal, trained with 40k sequence length and sliding-window attention for improved long-context performance. It is mirrored across the following Hugging Face repositories:

• z-lab/Qwen3.5-9B-DFlash
• modal-labs/Qwen3.5-9B-DFlash

This repository contains a DFlash draft model for Qwen/Qwen3.5-9B. It is not a standalone language model. It is intended to be paired with the target model in a speculative decoding server.

DFlash uses a lightweight block diffusion draft model to propose multiple tokens in parallel. The target model verifies those proposals, improving serving throughput while preserving the target model's output distribution.

Quick Start

This model should be used with an inference server that supports DFlash speculative decoding. An example SGLang deployment is:

export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.5-9B \
  --trust-remote-code \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path modal-labs/Qwen3.5-9B-DFlash \
  --speculative-dflash-block-size 8 \
  --speculative-draft-attention-backend fa4 \
  --attention-backend trtllm_mha \
  --linear-attn-prefill-backend flashinfer \
  --linear-attn-decode-backend flashinfer \
  --mamba-scheduler-strategy extra_buffer \
  --tp-size 1 \
  --max-running-requests 32 \
  --cuda-graph-max-bs-decode 32 \
  --cuda-graph-backend-prefill tc_piecewise \
  --enable-flashinfer-allreduce-fusion \
  --mem-fraction-static 0.8 \
  --host 0.0.0.0 \
  --port 30000

Block size 8 is the recommended default for higher-concurrency serving. Block size 16 gives longer accept lengths and strong concurrency-1 throughput in most workloads.

Benchmark Results

We benchmarked DFlash against the autoregressive baseline and Qwen's built-in MTP draft path. DFlash reaches up to 5.01x speedup at concurrency 1 and 2.58x at concurrency 32. Across the benchmark suite, DFlash delivers higher throughput than MTP at every matched setting where both completed.

Setup

• Runtime: SGLang on 1x NVIDIA B200 GPU, tensor parallel size 1, bfloat16
• Backends: trtllm_mha target attention, fa4 DFlash draft attention, flashinfer linear-attention prefill and decode
• Workloads: GSM8K, MATH500, HumanEval, MBPP, and MT-Bench with the Qwen chat template
• Decoding: greedy, thinking enabled, max output length 4096 tokens
• Measurement: 5 independent runs per configuration at concurrency 1 and 32 with continuous batching
• Throughput: generated output tokens / wall-clock benchmark time, including prefill and scheduling
• Accept length: completion_tokens / spec_verify_ct per generation turn, averaged across generation turns

Throughput and Speedup

Each cell is output tok/s (speedup). Bold marks the fastest speculative configuration in each row.

Concurrency 1

Concurrency 32

Accept Length

Mean accept length at concurrency 1. Bold marks the higher value in each matched MTP/DFlash pair.

Citation

If you find DFlash useful, please cite the original paper:

@article{chen2026dflash,
  title   = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026}
}