VideoDeltaNet open-sources hybrid attention that speeds up MiniMax H3 video generation up to 90x
realmrfakename · x · 2026-09-03
OpenVDN released VDN-H3, a hybrid-attention video model built on MiniMax H3 claiming up to 90x speedup with near-lossless quality, outperforming FastH3.
- Hybrid architecture: a frame-wise linear attention branch provides efficiency while a softmax branch preserves backbone visual quality and consistency.
- Plug-and-play: implemented as a new linear attention branch plus two small LoRA adapters merged at inference, leaving backbone weights untouched.
- Speed: generates a 14.4-second clip in 11.23 seconds on 8x B200 with 8 denoising steps — faster than real-time playback.
- Fully open source: 82 GB of weights (including the 50-step distillation source and 8-step turbo model), an optimized inference stack, and training code are all released.
Related event: Open-Source Video Delta Net Pushes Video Generation Faster Than Playback(8 posts)→
More from Infra
- PyTorch Foundation ecosystem grows 2.4x contributors across six projects including vLLM and Ray — PyTorch · 2026-10-21
- TorchTPU launches today — Rob Mulla publishes explainer video on how it works — Rob_Mulla · 2026-10-21
- DeepMind infra lead calls Gartner's 'compute surplus' claim nonsense: we're massively constrained — appenz · 2026-10-21
- FastVideo open-sources MiniMax FastH3 Omni with training code, 12x faster on Blackwell — multimodalart · 2026-10-21
- Anthropic's draft IPO filing: $16.3B H1 revenue, $80B net loss, $750B compute plans — kimmonismus · 2026-10-21
- dphnAI's distributed inference network served 32B DeepSeek tokens in 24h, ~6% of Venice's daily volume — Ronangmi · 2026-10-21