Video DeltaNet Renders a 14.3-Second 768p Clip in 6.70 Seconds on Eight B200s, and SGLang Already Ships It
Video DeltaNet (arXiv 2609.20744, submitted 17 Sep 2026) combines local Softmax attention with a bidirectional linear memory branch whose Video Delta Attention updates memory once per frame across that frame's spatial tokens, using separate output projections, learnable gates and a staged teacher-alignment recipe to graft the pathway onto pretrained models. Instantiated on MiniMax H3 with the hybrid applied only to video-to-video interactions and Softmax retained for text and audio, eight-step distillation plus an optimized SGLang serving stack completes DiT denoising for a 14.3-second 768p video in 6.70 seconds on eight NVIDIA B200 GPUs, a 14.5x speedup over the 50-step dense H3 baseline on the same GPUs. SGLang v0.5.20, released 18 September 2026, lists VDN-H3 among its supported diffusion models, so this is shipped rather than paper-only.
↳ Follow the thread