DFLASH), no patches needed (tested with SGLang 0.5.23 and vLLM 0.33.0).| Draft | Target model | Draft size | Block size |
|---|---|---|---|
| Qwen3.5-4B-MMFlash | Qwen/Qwen3.5-4B | 537M | 16 |
| Qwen3.5-9B-MMFlash | Qwen/Qwen3.5-9B | 1.05B | 16 |
| Qwen3.8-27B-MMFlash | Qwen/Qwen3.8-27B | 1.73B | 16 |
| Gemma4-E4B-it-MMFlash | google/gemma-4-E4B-it | 537M | 16 |
@article{feng2026mmflash,
title={MMFlash: Faster Speculative Decoding for Vision-Language Models},
author={Feng, Sicheng and Chen, Zigeng and Wang, Song and Fang, Gongfan and Wang, Xinchao},
journal={arXiv preprint arXiv:2611.18482},
year={2026},
}