GLM-5.3-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for zai-org/GLM-5.3. 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"

sglang serve \
  --model-path zai-org/GLM-5.3 \
  --tp-size 4 \
  --trust-remote-code \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/GLM-5.3-DFlash2 \
  --speculative-draft-attention-backend fa4

DFlash 2 is also supported by vLLM v0.28.0 and later; see incoai/GLM-5.3-NVFP4 for a vLLM serving example with the NVFP4-quantized target. See the blog post for more details.

Evaluation

• Runtime: SGLang on four NVIDIA GB300 GPUs (TP4), with FlashAttention 4 for DFlash 2 draft attention
• Speculation block size: 8 (7 draft tokens per verification step)
• Sampling: GLM-5.3's officially recommended parameters (temperature 1.0, top-p 0.95), with the default Max reasoning effort
• Maximum new tokens: 4096
• Samples: 128 at concurrency 1; 1,024 at concurrency 8 and 32

We compare autoregressive decoding, GLM-5.3's native MTP, 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	DFlash 2
GSM8K	5.12	5.94
MATH-500	5.05	6.02
HumanEval	4.85	5.48
MBPP	4.34	4.95
MT-Bench	3.81	4.19

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	113.4	292.6 (2.58×)	366.6 (3.23×)
MATH-500	113.1	297.4 (2.63×)	383.3 (3.39×)
HumanEval	113.7	292.9 (2.58×)	363.7 (3.20×)
MBPP	113.4	266.5 (2.35×)	336.5 (2.97×)
MT-Bench	113.2	206.8 (1.83×)	244.3 (2.16×)

Concurrency 8

Task	Autoregressive	MTP	DFlash 2
GSM8K	535.3	1,094.5 (2.04×)	1,310.4 (2.45×)
MATH-500	549.6	1,145.6 (2.08×)	1,409.7 (2.56×)
HumanEval	554.4	1,133.8 (2.05×)	1,360.6 (2.45×)
MBPP	554.2	1,049.7 (1.89×)	1,277.4 (2.31×)
MT-Bench	544.5	807.0 (1.48×)	895.5 (1.64×)

Concurrency 32

Task	Autoregressive	MTP	DFlash 2
GSM8K	1,142.7	2,283.6 (2.00×)	2,694.5 (2.36×)
MATH-500	1,251.8	2,943.2 (2.35×)	3,559.8 (2.84×)
HumanEval	1,303.1	3,016.9 (2.32×)	3,589.1 (2.75×)
MBPP	1,292.8	2,790.2 (2.16×)	3,380.4 (2.61×)
MT-Bench	1,262.7	2,119.5 (1.68×)	2,345.0 (1.86×)

License

This model is released under CC BY-NC-ND 4.0 for research and evaluation. For commercial licensing, contact contact@inco.ai.

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}
}