MMFlash from NUS speeds up vision-language speculative decoding up to 3.58x

NationalUniversityofSingapore · hf · 2026-11-16

Researchers at the National University of Singapore introduce MMFlash, improving multimodal speculative decoding by adaptively upweighting visually dependent tokens (identified via a KL-divergence-based dependence metric) and using sparse attention for long video. It achieves up to 3.58x lossless speedup across text, STEM, and video tasks, with code and models open-sourced.

Original post →

More from Research

Research channel →