FastVideo open-sources FastH3 Omni: 12x faster generation on NVIDIA Blackwell, training code included

Loading video

Loading video

Hao AI Lab's FastVideo team open-sources FastH3 Omni, an accelerated build of MiniMax H3. One request can mix up to 9 images plus video clips and voice references, each addressed in the prompt as <Picture 1> / <Video 1> / <Audio 1> so you pick whose face, which motion and whose voice goes where. PDD + DMD distillation cuts end-to-end latency 12x versus base MiniMax-H3 Ref2VA at 50 steps, measured over 27 matched workloads on 8x GB200 with identical prompts, references and 1344x768 output, timed until the MP4 is written. NVFP4/FP8 weights target DGX Spark and RTX, ComfyUI is supported day 0, and the training code plus reference-aware preprocessing are public.