FastH3

Updated

FastH3 is an optimized, distilled derivative of the MiniMax H3 model, developed by the FastVideo team (affiliated with UCSD's Hao AI Lab). It is designed to accelerate synchronized video-and-audio generation by significantly reducing the computational requirements of the original model.

Key Features and Performance

  • Efficiency: FastH3 reduces the number of required inference transformer calls from 49 (in the base MiniMax H3) to 4. Additionally, it employs Video Sparse Attention (VSA) to prune approximately 90% of the computational load per step.
  • Speed: These optimizations result in up to a 14x increase in generation speed on a single NVIDIA Blackwell GPU. On a system with eight NVIDIA B200 GPUs, it can generate a 15-second, 768p video in less than 13 seconds, making it faster than real-time playback.
  • Functionality: It maintains the core capability of the MiniMax H3 model, allowing users to generate complete short-form videos with synchronized visuals, dialogue, music, and ambient sound from a single text prompt.
  • Flexibility: The model supports variable durations, resolutions, and aspect ratios (including square, portrait, landscape, and ultrawide) within a single checkpoint.

Availability and Licensing

  • Open Weights: FastH3 is considered an "open-weight" model. While it uses MiniMax H3 weights—which are subject to the MiniMax H3 Community License—the FastVideo inference framework itself is Apache-2.0 open source.
  • Deployment: The model can be run on various hardware, including data-center NVIDIA GPUs and local paths optimized for NVIDIA DGX Spark and Apple Silicon. It is also integrated into several third-party platforms as a "Turbo" engine for faster content creation.