HyperFlow H3 Flash shrinks a huge video model to run on consumer GPUs

A project called Video Rebirth has released HyperFlow H3 Flash, a compressed version of the large MiniMax-H3 video generation transformer. The original model has about 31 billion parameters, but this new version cuts that down to 13.7 billion parameters for generating video with audio from a first and last frame, or 13.9 billion parameters when using reference images to guide generation. The model can produce results in as few as 4, 6, or 8 sampling steps, and it does not need classifier-free guidance, a technique that normally adds extra computation. In practical terms, generating a 5.2-second video clip using 6 steps takes just 8.7 seconds on four B200 GPUs, 36 seconds on a single H200 GPU, and 92 seconds on a single RTX 5090, a GPU many enthusiasts and small studios already own. The compressed model is also available in a smaller file format called Q4_K_M GGUF, which is only 9.1 gigabytes, making it practical to download and run locally without specialized data center hardware. This kind of compression work helps bring advanced video generation capabilities to people without access to expensive cloud computing.

Why it matters

Shrinking a 31-billion parameter video model down to run on a single consumer GPU makes advanced video generation accessible to far more people.

Source: Video Rebirth