AnalysisAI ModelsNovember 14, 2026

MMFlash speeds vision-language decoding up to 3.58×

Read original source →arxiv.org

MMFlash adaptively upweights visually dependent tokens and adds sparse attention for long video, yielding up to 3.58× lossless speculative decoding speedup. It targets multimodal drafting, which diffusion drafters like DFlash largely left to text-only tasks.

2 sources

More stories today

Open the live feed