MMFlash boosts vision-language speculative decoding with up to 3.58x lossless speedup
_akhaliq · x · 2026-11-16
A new Hugging Face paper introduces MMFlash, a speculative decoding framework tailored to vision-language models.
- The authors propose a token-level visual dependence metric based on KL divergence, finding acceptance rates drop from 46% (lowest visual-dependence quartile) to 22% (highest), a gap persisting even after multimodal training
- MMFlash adaptively upweights visually dependent tokens that the diffusion drafter predicts poorly, and uses sparse attention for long video inputs to cut redundant computation and improve drafting quality
- Across multimodal benchmarks it achieves lossless speedups up to 3.58x, beating EAGLE-3 by up to 195.4% and DFlash by 17.1%
More from Infra
- LateTalk guest: AI infrastructure buildout may already be done, apps are still early — oran_ge · 2026-11-16
- Export controls target China's EUV self-sufficiency, not GPU smuggling, analyst argues — teortaxesTex · 2026-11-16
- Cyberpunk 2077 gets a stunning visual overhaul with DLSS 5 Double Pass and Path Tracing — Promptmethus · 2026-11-16
- Microsoft open-sources TokenEconomizer, cutting coding agent costs 18-31% — udmrzn · 2026-11-16
- Engy (SN53) out-processes Venice in tokens with ~28x smaller market cap — markjeffrey · 2026-11-16
- US chipmaking capacity fell 37%→10%; open-weights optimism mirrors factory reshoring talk — ccerrato147 · 2026-11-16