Qwen DFlash2

Block-diffusion draft · z-lab · Runtime companion

z-lab's Qwen3.8 27B DFlash2 draft model, containing its config and safetensors weights. It drafts tokens for a target model to verify and is not a standalone language model.

Benchmark results · Files · Download · Runtime · Provenance · Full collection

Join Infinity on Discord

Get model support, request models and quants, discuss benchmarks, and follow new releases and model improvements.

Join the Infinity community

All model tests, per-question scores, experiment history and scoring methods

Qwen effort study — separate full-404 cohort: medium 82.87/100 versus xhigh 84.41/100; paired difference +1.54 points (95% interval [-1.71, +4.82]). 404 collected questions, 403 valid questions graded per effort; one invalid reference annotation excluded symmetrically. The historical 202-question cross-model comparison below is unchanged. Study, timing and blind subjective results.

Preserved and documented by isilicon / Mango. The original model, modification and quantization remain the upstream creators' work. This repository adds attributed preservation and exact-configuration evaluation, with the selected source bytes unchanged.

At a glance

Property	Recorded value
Artifact	safetensors draft
Selected source payload	3.85 GB / 3.58 GiB
Primary selected files	2
Evaluation status	Runtime companion
Source publisher	z-lab
Source revision	`50307d4c4cde6860d4eee73e2547cd786fe8e8a4`
Declared license	apache-2.0

GB is decimal and GiB is binary. Payload size is disk/storage size, not a claim that the model fits in the same amount of VRAM. Runtime, KV cache, working memory and companions add their own requirements.

Mango evaluation

No standalone Mango quality score has been published for these exact files. This is a companion artifact; the consuming model's result belongs to the full target/runtime configuration.

Related configurations measured separately

These results describe the linked configurations, not this artifact. A change of quant, modification lineage, draft, engine or tokenizer can change the result.

Different configuration	Its weighted score	Its workload min
Qwen3.8 27B Cyber · IQ4_XS imatrix · embedded MTP	70.27	29.74
Qwen3.8 27B Cyber · Q6_K · Q8 MTP + projector	66.98	45.60
Qwen3.8 27B · MLX affine 4-bit · OrcaRouter	80.64	23.98

A completed score requires a native preflight, matching frozen prompts, 202 suite answers, 12 agent cases, archive verification and the same reference graders. The benchmark methodology describes the protocol; publisher-side results are retained separately below.

Runtime and compatibility

A speculative-decoding runtime must support DFlash2 and the target model. The measured Mango workload combines the target, draft and engine; this mirror alone has no separate task-quality or speed score.

Context and output behavior belong to the target/runtime configuration.

Companion artifacts

Consuming model	Connection
Qwen3.8 27B · MLX affine 4-bit · OrcaRouter	Use the compatible parent runtime and exact target revision

Files and integrity

Selected file · original path	GB	Identity prefix
config.json	0.00000	`79279cc5665bced6…` (Git blob)
model.safetensors	3.84882	`67fc76d68dc5a941…` (SHA-256)

Full paths, byte sizes and complete SHA-256/Git blob identities are in SOURCE_MANIFEST.json. Available original README/license/notice files are also preserved. MIRROR_STATUS.json records the earlier source-file verification; MODEL_CARD_STATUS.json records this documentation refresh separately.

Download and use

The commands pin the already-verified mirror weight revision. Documentation can advance on main without changing that preserved weight snapshot. Keep original subfolders and all shards.

python -m pip install -U huggingface_hub

hf download isilicon/mango-dflash2 \
  config.json \
  model.safetensors \
  --revision e383e96a1ac8822a9012a57d43c40d488993a7ad \
  --local-dir ./weights/dflash2

Load example and setup scope

This component has no independent chat launch command. Install it through the consuming model's compatible runtime and companion layout.

Provenance and license

The source credits the DFlash2 work and points to the incoai release. Its acceptance-length and SGLang/vLLM throughput tables concern the publisher's configurations. Mango used this companion with its Qwen MLX4 target through TensorFold.

Responsibility / identity	Attribution
Mirror / local evaluation	isilicon · Mango
Immediate source	z-lab/Qwen3.8-27B-DFlash2
Pinned source commit	`50307d4c4cde6860d4eee73e2547cd786fe8e8a4`
Declared upstream base(s)	Qwen/Qwen3.8-27B
Source-declared license	apache-2.0

Original notices, base-model conditions and the publisher's stated limitations still apply. Downloadable preservation does not add new license rights. Source-described refusal modification and publisher-side benchmark claims have not been independently re-measured by this Mango task cohort.

Original publisher documentation

The unchanged upstream README and pinned source card retain the original usage guidance, technical discussion, attribution, caveats and citations.

Read the complete original publisher card

Source reference. The following is the original publisher's documentation. Its other quants, full-model files, context claims and benchmark tables can exceed the selection in this mirror. Relative asset/file URLs are qualified to the pinned original source; publisher measurements remain separate from Mango measurements.

Qwen3.8-27B-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for Qwen/Qwen3.8-27B. It is not a standalone language model: it runs inside a speculative decoding server and drafts tokens for the target model to verify. This repository is a mirror of incoai/Qwen3.8-27B-DFlash2.

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.

DFlash 2: parallel block drafting with a candidate path selector

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/Qwen3.8-27B \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Qwen3.8-27B-DFlash2 \
  --speculative-num-draft-tokens 8

Or with vLLM:

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

vllm serve Qwen/Qwen3.8-27B \
  --speculative-config '{
    "method": "dflash",
    "model": "incoai/Qwen3.8-27B-DFlash2",
    "num_speculative_tokens": 7
  }'

See the blog post for other engines and more details.

Evaluation

Runtime: SGLang on one NVIDIA H200, with FlashAttention 3 for target and draft attention
Speculation block size: 8 (7 draft tokens per verification step)
Sampling: Qwen3.8's officially recommended parameters (temperature 1.0, top-p 0.95, top-k 20), with xhigh reasoning effort
Maximum new tokens: 4096
Prompts: benchmark formatting from z-lab/dflash

We compare autoregressive decoding, Qwen3.8's built-in seven-token MTP, a community DSpark drafter (RadixArk/Qwen3.8-27B-DSpark), and DFlash 2. All 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. Higher is better.

Task	MTP	DSpark	DFlash 2
GSM8K	5.02	4.36	5.46
MATH-500	4.72	3.92	5.28
HumanEval	3.91	3.30	4.39
MBPP	3.99	3.51	4.79
MT-Bench	3.74	3.01	4.10

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	DSpark	DFlash 2
GSM8K	68.9	178.5 (2.59×)	185.3 (2.69×)	236.1 (3.43×)
MATH-500	69.0	172.8 (2.51×)	174.5 (2.53×)	230.7 (3.34×)
HumanEval	69.0	151.9 (2.20×)	159.9 (2.32×)	214.6 (3.11×)
MBPP	69.0	153.1 (2.22×)	163.3 (2.37×)	226.9 (3.29×)
MT-Bench	68.9	134.9 (1.96×)	137.6 (2.00×)	184.0 (2.67×)

Concurrency 8

Task	Autoregressive	MTP	DSpark	DFlash 2
GSM8K	467.2	1,022.1 (2.19×)	1,040.8 (2.23×)	1,328.7 (2.84×)
MATH-500	480.0	1,023.5 (2.13×)	1,025.8 (2.14×)	1,368.3 (2.85×)
HumanEval	483.4	934.2 (1.93×)	956.5 (1.98×)	1,291.5 (2.67×)
MBPP	478.0	938.1 (1.96×)	974.1 (2.04×)	1,328.0 (2.78×)
MT-Bench	480.5	835.2 (1.74×)	802.3 (1.67×)	1,090.2 (2.27×)

Concurrency 32

Task	Autoregressive	MTP	DSpark	DFlash 2
GSM8K	1,329.8	1,381.1 (1.04×)	1,506.5 (1.13×)	1,922.5 (1.45×)
MATH-500	1,505.8	1,415.6 (0.94×)	1,429.0 (0.95×)	1,951.8 (1.30×)
HumanEval	1,546.5	1,296.8 (0.84×)	1,330.1 (0.86×)	1,799.0 (1.16×)
MBPP	1,507.7	1,314.9 (0.87×)	1,361.3 (0.90×)	1,886.8 (1.25×)
MT-Bench	1,507.4	1,159.7 (0.77×)	1,115.5 (0.74×)	1,525.3 (1.01×)

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}
}

Preservation record

Selected source bytes are verified against pinned metadata; the original eight graded mirrors also passed the previously recorded anonymous 1 MiB weight-prefix checks. This is not a full independent-device restore/load test. Hugging Face repository copies share a provider and do not guarantee permanent storage.

Collection · Source manifest · Original documentation