MMFlash speeds up vision-language models up to 3.58x with lossless speculative decoding

NUS researchers introduced MMFlash, a vision-aware speculative decoding framework for vision-language models that uses KL-divergence-based token-level visual dependency weighting and sparse attention for long videos, achieving up to 3.58x lossless inference speedup.

2026-11-16 ~ 2026-11-16 · 2 related posts