Sparse attention with Jev cuts MiniMax H3 video generation time by 41.7% on an RTX 4070

TheMoonMidas · x · 2026-09-20

A Japanese developer integrated the trending Jev scheme into MiniMax H3's video generation pipeline to sparsify attention processing: Jev judges importance per layer (49 layers across 4 steps) and dynamically selects sparsity rates of 1%, 3%, 5%, or 10%. On an RTX 4070, generation time dropped from 6:07 to 3:34 — a 41.7% speedup — even while the pipeline queries the Jev cloud during generation.

Original post →

More from Multimodal

Multimodal channel →