Qwen3.5-35B-A3B-DFlash

Paper | GitHub | Blog

DFlash is a speculative decoding method that uses a lightweight block diffusion model to draft multiple tokens in parallel, achieving up to 2.8x speedup over autoregressive decoding. This is the drafter model, which must be paired with Qwen/Qwen3.5-35B-A3B.

Quick Start

Installation

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

Launch Server

Use --speculative-num-draft-tokens to set the block size (8 or 16).

python -m sglang.launch_server \
    --model-path Qwen/Qwen3.5-35B-A3B \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3.5-35B-A3B-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/Qwen3.5-35B-A3B",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=4096,
)
print(response.choices[0].message.content)

vLLM

Community-contributed support is available. See PRs #36847 and #36767 for details.

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 2.8x speedup at concurrency 1.

Tokens/sec (speedup vs. autoregressive baseline)

Block Size = 16

Task	Concurrency	AR	MTP	DFlash
Math500	1	246	420 (1.7x)	681 (2.8x)
	8	1283	2165 (1.7x)	3108 (2.4x)
	16	1964	3393 (1.7x)	4721 (2.4x)
	32	2744	5042 (1.8x)	6637 (2.4x)
GSM8K	1	244	371 (1.5x)	594 (2.4x)
	8	1277	1923 (1.5x)	2672 (2.1x)
	16	1927	3000 (1.6x)	3965 (2.1x)
	32	2678	4368 (1.6x)	5483 (2.0x)
HumanEval	1	237	404 (1.7x)	662 (2.8x)
	8	1206	1906 (1.6x)	2721 (2.3x)
	16	1796	2958 (1.6x)	3899 (2.2x)
	32	2433	4027 (1.7x)	5066 (2.1x)
MBPP	1	240	356 (1.5x)	617 (2.6x)
	8	1228	1753 (1.4x)	2547 (2.1x)
	16	1854	2608 (1.4x)	3591 (1.9x)
	32	2487	3730 (1.5x)	4785 (1.9x)
MT-Bench	1	244	307 (1.3x)	489 (2.0x)
	8	1273	1545 (1.2x)	2198 (1.7x)
	16	1901	2389 (1.3x)	3213 (1.7x)
	32	2657	3513 (1.3x)	4327 (1.6x)
Alpaca	1	245	287 (1.2x)	456 (1.9x)
	8	1276	1468 (1.2x)	2067 (1.6x)
	16	1917	2276 (1.2x)	3059 (1.6x)
	32	2667	3373 (1.3x)	4224 (1.6x)

Block Size = 8

Task	Concurrency	AR	MTP	DFlash
Math500	1	246	514 (2.1x)	621 (2.5x)
	8	1283	2440 (1.9x)	2811 (2.2x)
	16	1964	3848 (2.0x)	4448 (2.3x)
	32	2744	5754 (2.1x)	6693 (2.4x)
GSM8K	1	244	469 (1.9x)	572 (2.3x)
	8	1277	2245 (1.8x)	2584 (2.0x)
	16	1927	3477 (1.8x)	4040 (2.1x)
	32	2678	5209 (1.9x)	5885 (2.2x)
HumanEval	1	237	496 (2.1x)	596 (2.5x)
	8	1206	2153 (1.8x)	2427 (2.0x)
	16	1796	3318 (1.8x)	3664 (2.0x)
	32	2433	4678 (1.9x)	5016 (2.1x)
MBPP	1	240	453 (1.9x)	583 (2.4x)
	8	1228	2060 (1.7x)	2441 (2.0x)
	16	1854	3096 (1.7x)	3605 (1.9x)
	32	2487	4403 (1.8x)	4998 (2.0x)
MT-Bench	1	244	406 (1.7x)	486 (2.0x)
	8	1273	1942 (1.5x)	2204 (1.7x)
	16	1901	2973 (1.6x)	3418 (1.8x)
	32	2657	4404 (1.7x)	4977 (1.9x)
Alpaca	1	245	396 (1.6x)	479 (2.0x)
	8	1276	1853 (1.5x)	2110 (1.7x)
	16	1917	2877 (1.5x)	3315 (1.7x)
	32	2667	4297 (1.6x)	4860 (1.8x)

Acceptance Length

Format: MTP / DFlash (average over concurrency 1, 8, 16, 32)

Task	B8	B16
Math500	5.59 / 5.51	6.93 / 7.20
GSM8K	5.39 / 5.27	6.58 / 6.78
HumanEval	5.62 / 5.69	7.14 / 7.90
MBPP	4.92 / 5.07	5.71 / 6.45
MT-Bench	4.49 / 4.41	5.20 / 5.35
Alpaca	4.32 / 4.23	4.87 / 5.04

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}
}