Qwen3.6-35B-A3B-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for Qwen/Qwen3.6-35B-A3B. It is not a standalone language model: it runs inside a speculative decoding server and drafts tokens for the target model to verify.

DFlash 2 is a block-diffusion drafter for speculative decoding. It predicts a whole block of tokens in a single pass and keeps the top candidates at every position. A lightweight selector then traces one coherent path through them. Two-tap dynamic convolutions in the backbone keep the draft from decaying toward the end of the block. Decoding is lossless: greedy output matches the target model exactly, and sampling preserves its distribution.

Quick Start

Serve with SGLang:

pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.6-35B-A3B \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Qwen3.6-35B-A3B-DFlash2 \
  --speculative-num-draft-tokens 8 \
  --attention-backend flashinfer \
  --speculative-draft-attention-backend fa4

Or with vLLM:

pip install -U "vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/52816/head"

vllm serve Qwen/Qwen3.6-35B-A3B \
  --speculative-config '{
    "method": "dflash",
    "model": "incoai/Qwen3.6-35B-A3B-DFlash2",
    "num_speculative_tokens": 7
  }'

See the blog post for other engines and more details.

Evaluation

Runtime: SGLang on one NVIDIA GB300, with FlashInfer for target attention and FlashAttention 4 for draft attention
Speculation block size: 8 (7 draft tokens per verification step)
Sampling: temperature 1.0, top-p 0.95, top-k 20, presence penalty 0.0, repetition penalty 1.0, with thinking enabled
Maximum new tokens: 4096
Requests: 32 per dataset and concurrency, with seed 42 for prompt selection
Prompts: benchmark formatting from z-lab/dflash; MT-Bench uses the first turn

We compare autoregressive decoding, Qwen3.6's built-in MTP with seven speculative steps, and DFlash 2. Both speculative methods propose seven draft tokens per verification step.

Acceptance Length

Acceptance length is the per-request mean of completion tokens divided by verification steps at concurrency 1. Higher is better.

Task	MTP	DFlash 2
GSM8K	4.92	5.21
MATH-500	5.02	5.34
HumanEval	4.44	4.90
MBPP	4.36	4.76
MT-Bench	4.26	4.41

Throughput

Throughput is total output tokens divided by end-to-end wall time. Each cell shows output tok/s (speedup vs. autoregressive).

Concurrency 1

Task	Autoregressive	MTP	DFlash 2
GSM8K	358.6	505.1 (1.41×)	670.0 (1.87×)
MATH-500	358.8	542.1 (1.51×)	749.4 (2.09×)
HumanEval	358.7	480.1 (1.34×)	680.3 (1.90×)
MBPP	361.0	474.2 (1.31×)	660.1 (1.83×)
MT-Bench	360.5	425.4 (1.18×)	550.0 (1.53×)

Concurrency 8

Task	Autoregressive	MTP	DFlash 2
GSM8K	1,524.0	1,850.0 (1.21×)	2,323.7 (1.52×)
MATH-500	1,600.0	2,391.2 (1.49×)	3,018.2 (1.89×)
HumanEval	1,595.9	2,126.1 (1.33×)	2,814.4 (1.76×)
MBPP	1,503.4	2,024.6 (1.35×)	2,532.5 (1.68×)
MT-Bench	1,460.6	1,727.4 (1.18×)	2,146.2 (1.47×)

Concurrency 32

Task	Autoregressive	MTP	DFlash 2
GSM8K	2,499.4	2,957.5 (1.18×)	3,971.2 (1.59×)
MATH-500	3,432.0	5,201.4 (1.52×)	6,483.1 (1.89×)
HumanEval	3,060.7	4,361.9 (1.43×)	5,621.9 (1.84×)
MBPP	2,777.0	3,700.0 (1.33×)	5,025.9 (1.81×)
MT-Bench	2,535.3	3,082.5 (1.22×)	3,705.5 (1.46×)

License

This model is released under the Apache License 2.0.

Citation

If you find DFlash 2 useful, please cite:

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}

Please also cite the original DFlash paper:

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}