FastVideo open-sources FastH3: 4-step distilled MiniMax H3 with 14x Blackwell speedup
victormustar · x · 2026-10-31
Responding to how expensive and slow video generation remains, UCSD's Hao AI Lab (FastVideo team), with NVIDIA FastGen and Nuva Lab, open-sourced FastH3 Preview v1 — an open-weight text-to-video-and-audio (T2VA) model post-trained on MiniMax H3.
- Speed: 15s of 768p video in under 13s on 8xB200; up to 14x speedup on a single Blackwell GPU
- Quality: 1k+ B200 training hours with real-world multi-shot, audio-synced distributions; one checkpoint natively supports variable resolution, aspect ratio and duration
- Openness: 4-step VSA/data-free checkpoints, full weights, pre-extracted LoRA, ablations, open inference code (training recipe coming)
- Roadmap: image-ref (FL2VA), FastH3 Omni, Nvfp4, optimizations for RTX/DGX Sparks/Apple MLX, and retraining with NVIDIA's Parallel Decoding Distillation
The team positions it as a foundation for real commercial workloads, not just academic experiments.
More from Multimodal
- AI-generated Doritos USB drive concept mocks design-product photography — Delahuntagram · 2026-10-31
- Krea 2 + H3 morph workflow auto-generates soundtracks via SigLIP2 and Suno MCP — Accurate_Public4674 · 2026-10-31
- Suno ships MCP Beta, dev uses it to sync motion graphics to AI-generated music — enggirlfriend · 2026-10-31
- Mischief launches video model trained from lab-grown brain cells at 0.5¢ per second — evilsocket · 2026-10-31
- A full structured prompt for a 30-second hyper-real horror video sequence — LudovicCreator · 2026-10-31
- Creepy AI horror short made with Seedance 2.5: the abandoned hospital hallway — LudovicCreator · 2026-10-31