Flint-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for RecursiveSI/flint. 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. Flint support is in an open pull request (#45624), so install it at the commit we validated:

git clone https://github.com/sgl-project/sglang.git
cd sglang
git fetch origin 37bc771d2b253e2d87d27ace71297154874ee3b4
git checkout 37bc771d2b253e2d87d27ace71297154874ee3b4
pip install -e "python[all]"

sglang serve RecursiveSI/flint \
  --enable-fp32-lm-head \
  --reasoning-parser flint \
  --tool-call-parser flint \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Flint-DFlash2

See the blog post for more details.

Evaluation

• Runtime: SGLang (#45624 at 37bc771d) on one NVIDIA GB300 (TP1), in BF16, with FlashAttention 4 for target and draft attention
• Speculation block size: 8 (7 draft tokens per verification step)
• Sampling: Flint's default parameters (temperature 1.0, top-p 0.95, top-k 20), with its default chat template and reasoning
• Maximum new tokens: 4096
• Samples: 128 at concurrency 1; 1,024 at concurrency 8 and 32
• Prompts: benchmark formatting and repeated sampling from z-lab/dflash

We compare autoregressive decoding and DFlash 2, which proposes 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	DFlash 2
GSM8K	4.23
MATH-500	4.28
HumanEval	4.53
MBPP	4.33
MT-Bench	3.49

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	DFlash 2
GSM8K	316.7	686.5 (2.17×)
MATH-500	316.7	651.5 (2.06×)
HumanEval	308.6	715.5 (2.32×)
MBPP	310.8	700.4 (2.25×)
MT-Bench	313.6	555.7 (1.77×)

Concurrency 8

Task	Autoregressive	DFlash 2
GSM8K	1,474.7	2,720.8 (1.84×)
MATH-500	1,500.2	2,651.3 (1.77×)
HumanEval	1,490.0	2,828.3 (1.90×)
MBPP	1,504.2	2,705.8 (1.80×)
MT-Bench	1,487.9	2,151.9 (1.45×)

Concurrency 32

Task	Autoregressive	DFlash 2
GSM8K	3,523.4	6,936.8 (1.97×)
MATH-500	3,534.8	6,662.9 (1.88×)
HumanEval	3,554.4	7,374.5 (2.07×)
MBPP	3,505.6	7,168.3 (2.04×)
MT-Bench	3,385.8	5,792.0 (1.71×)

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}
}