VideoDeltaNet: Near-Lossless Hybrid Attention on MiniMax H3
Loading video
Loading videoVideoDeltaNet delivers near-lossless hybrid attention on MiniMax H3. Local softmax preserves detail while linear attention handles long range, so a 14.4s 768p clip generates in 11.23s on 8 B200s: faster than you can watch it, streamed continuously, with open weights.
MiniMax H3VideoDeltaNetB200Video GenerationHybrid AttentionLinear AttentionWorld ModelsOpen Source Weights
Category: research
Author: @wildmindai
Date: 2026-09-03T00:00:00
Duration: 53.717s
