Fastvideo · v3 · 2× · naposledy 03. 9. 2026
FastH3 (Preview v1) is an Open Source AI model for text-to-video-and-audio generation developed by FastVideo (Hao AI Lab, UC San Diego) in collaboration with Nuva Lab and the NVIDIA-FastGen team. It is a 4-step DMD2 Distillation of the 33B-parameter model MiniMax H3, reducing 49 Transformer calls of the base model to 4 and additionally utilizing 90% sparse Attention, achieving up to 14x acceleration on a single NVIDIA Blackwell GPU. It generates synchronized video (up to 768p, 24 FPS) and stereo audio in a single Inference pass; via vLLM-Omni, a 10.125-second MP4 output was demonstrated in under 9 seconds on 8x B300 GPUs. FastH3 inherits the MiniMax H3 Community License, which restricts commercial use and [text appears incomplete].