FastH3 open-weight preview cuts MiniMax H3 video and audio generation to four DiT calls
- Hao AI Lab, Nuva Lab, and NVIDIA FastGen released FastH3 Preview v1, an open-weight distilled version of MiniMax H3 with model weights and a pre-extracted LoRA.
- The team reports up to a 14x speedup on one NVIDIA Blackwell GPU and generation of a 15-second, 768p video in under 13 seconds across eight NVIDIA B200 GPUs.
- The recommended checkpoint reduces MiniMax H3 inference to four DiT calls using DMD2-based few-step distillation, rather than the original model's substantially higher number of transformer evaluations.
- FastH3 uses Video Sparse Attention at 90% sparsity with a tile size of 64, and reported performance requires the FastVideo VSA-H3 backend and kernel.
- The preview retains native text-to-video-and-audio generation and reuses H3's text encoder, video and audio VAEs, tokenizers, schedulers, and base transformer architecture.