Qwen3.5-9B-DFlash

DFlash is a speculative decoding method that uses a lightweight block diffusion model to draft multiple tokens in parallel, achieving up to 4.4x speedup over autoregressive decoding. This is the drafter model, which must be paired with Qwen/Qwen3.5-9B.

Quick Start

Installation

vLLM:

uv pip install vllm
uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly

SGLang:

uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/20547/head#subdirectory=python"

Launch Server

vLLM:

vllm serve Qwen/Qwen3.5-9B \
  --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.5-9B-DFlash", "num_speculative_tokens": 15}' \
  --attention-backend flash_attn \
  --max-num-batched-tokens 32768

SGLang:

# Optional: enable schedule overlapping (experimental, may not be stable)
# export SGLANG_ENABLE_SPEC_V2=1
# export SGLANG_ENABLE_DFLASH_SPEC_V2=1
# export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

python -m sglang.launch_server \
    --model-path Qwen/Qwen3.5-9B \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3.5-9B-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

Tip: For long-context or agentic workloads, add --speculative-dflash-draft-window-size WINDOW_SIZE to enable sliding-window attention for the drafter.

Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=4096,
    temperature=0.0
)
print(response.choices[0].message.content)

Benchmark Results

Setup: Single NVIDIA B200, SGLang, thinking enabled, max output length 4096. We report end-to-end throughput, including prefill time. See our GitHub repository for reproduction scripts.

Throughput and Speedup

DFlash outperforms MTP across all block sizes and concurrency levels, achieving up to 4.4x speedup at concurrency 1.

Tokens/sec (speedup vs. autoregressive baseline)

Block Size = 16

Task	Concurrency	AR	MTP	DFlash
Math500	1	197	379 (1.9x)	808 (4.1x)
	8	1472	2569 (1.7x)	5114 (3.5x)
	16	2831	4206 (1.5x)	7508 (2.7x)
	32	4701	6028 (1.3x)	9286 (2.0x)
GSM8K	1	198	342 (1.7x)	697 (3.5x)
	8	1470	2331 (1.6x)	4351 (3.0x)
	16	2781	3794 (1.4x)	6325 (2.3x)
	32	4581	5445 (1.2x)	7559 (1.6x)
HumanEval	1	193	378 (2.0x)	840 (4.4x)
	8	1414	2461 (1.7x)	4837 (3.4x)
	16	2638	3916 (1.5x)	6722 (2.5x)
	32	4217	5423 (1.3x)	8285 (2.0x)
MBPP	1	194	335 (1.7x)	755 (3.9x)
	8	1421	2064 (1.5x)	4202 (3.0x)
	16	2667	3358 (1.3x)	5843 (2.2x)
	32	4160	4610 (1.1x)	6961 (1.7x)
MT-Bench	1	194	297 (1.5x)	587 (3.0x)
	8	1451	1945 (1.3x)	3611 (2.5x)
	16	2787	3115 (1.1x)	5185 (1.9x)
	32	4578	4453 (1.0x)	6225 (1.4x)
Alpaca	1	197	278 (1.4x)	545 (2.8x)
	8	1460	1816 (1.2x)	3382 (2.3x)
	16	2789	3009 (1.1x)	5002 (1.8x)
	32	4574	4326 (1.0x)	6247 (1.4x)

Block Size = 8

Task	Concurrency	AR	MTP	DFlash
Math500	1	195	452 (2.3x)	664 (3.4x)
	8	1458	3199 (2.2x)	4703 (3.2x)
	16	2825	5390 (1.9x)	7804 (2.8x)
	32	4712	7941 (1.7x)	11003 (2.3x)
GSM8K	1	196	421 (2.1x)	591 (3.0x)
	8	1464	2954 (2.0x)	4106 (2.8x)
	16	2775	4939 (1.8x)	6733 (2.4x)
	32	4567	7246 (1.6x)	9375 (2.1x)
HumanEval	1	193	446 (2.3x)	667 (3.5x)
	8	1411	3020 (2.1x)	4366 (3.1x)
	16	2631	4884 (1.9x)	6815 (2.6x)
	32	4077	6819 (1.7x)	8899 (2.2x)
MBPP	1	197	409 (2.1x)	634 (3.2x)
	8	1440	2710 (1.9x)	3992 (2.8x)
	16	2682	4435 (1.7x)	6128 (2.3x)
	32	4152	6213 (1.5x)	8026 (1.9x)
MT-Bench	1	198	374 (1.9x)	525 (2.7x)
	8	1478	2612 (1.8x)	3668 (2.5x)
	16	2836	4323 (1.5x)	5905 (2.1x)
	32	4617	6335 (1.4x)	8288 (1.8x)
Alpaca	1	196	360 (1.8x)	503 (2.6x)
	8	1450	2497 (1.7x)	3493 (2.4x)
	16	2802	4194 (1.5x)	5714 (2.0x)
	32	4572	6175 (1.4x)	8077 (1.8x)

Acceptance Length

Format: MTP / DFlash

Task	B8	B16
Math500	5.46 / 5.67	6.66 / 7.34
GSM8K	5.27 / 5.33	6.37 / 6.71
HumanEval	5.39 / 5.87	6.61 / 7.93
MBPP	4.78 / 5.31	5.49 / 6.62
MT-Bench	4.52 / 4.53	5.30 / 5.49
Alpaca	4.38 / 4.35	5.03 / 5.10

Acknowledgements

Special thanks to David Wang for his outstanding engineering support on this project. We are also grateful to Modal, InnoMatrix, and Yotta Labs for providing the compute resources used to train this draft model.

Citation

If you find DFlash useful, please cite our work. To share feedback on DFlash or request new model support, please fill out this form: DFlash Feedback.

@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}
}

z-lab/Qwen3.5-9B-DFlash