HyperFlow H3 Flash - Few-step video and audio generation on consumer GPUs
Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8
HyperFlow H3 Flash from Video Rebirth compresses the ~31B MiniMax-H3 model down to 13.7B, generating video with synchronized stereo audio in just 4 to 8 sampling steps and no classifier-free guidance. It produces 5.2–15 second clips at 24 fps with a 32 kHz soundtrack, running on 24GB and 32GB consumer GPUs with only 16GB of system RAM thanks to GGUF quantization, or on data-center GPUs via INT8. With ComfyUI nodes, a Python API, and a CLI, it brings high-quality joint video-audio generation within reach of everyday creators.
Less than half the size of the original MiniMax-H3 transformer, yet it still delivers synchronized video and stereo audio in as few as six steps—running comfortably on a 24GB RTX GPU.