Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8

HyperFlow H3 Flash is a compressed version of MiniMax-H3, a roughly 31-billion-parameter model that generates video and synchronized stereo audio. The release reduces the transformer to 13.7 billion parameters for most tasks and supports generation in four, six, or eight sampling steps, without classifier-free guidance. It handles text-to-video, video from one or two keyframes, and reference-guided generation using images, video, and audio. Typical output is 24-fps video, 5.2 to 15 seconds long, with 32 kHz stereo sound; the default resolution is 1344×768. Reported six-step generation times for a 5.2-second clip range from 8.7 seconds on four B200 GPUs to 92 seconds on an RTX 5090.

Two deployment routes target different hardware. GGUF quantized weights are intended for consumer systems, including 24 GB GPUs and PCs with 16 GB system RAM; the recommended Q4_K_M version occupies about 9 GB, while Q8_0 uses about 15 GB. The Python runtime uses W8A8 INT8 weights and is aimed at data-center GPUs or high-memory RTX systems, with multi-GPU support. It requires a large first download - about 90 GB, or 68 GB with an FP8 text encoder - and compiles kernels on first use. ComfyUI, Unsloth Studio, and stable-diffusion.cpp integrations provide alternative workflows. The project’s central contribution is combining model compression, low-step sampling, and multiple runtimes to make joint audiovisual generation faster and less memory-intensive.