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

Researchers at Video Rebirth have compressed MiniMax-H3, a 31B parameter video and audio diffusion model, down to 14B parameters while reducing inference steps to just 6, achieving significant speed improvements and lower memory requirements. The resulting model, HyperFlow H3 Flash, can generate 5.2-15 seconds of synchronized video and audio and runs on consumer GPUs with as little as 24GB of VRAM, with optional quantization versions as small as 9.1GB. The model supports multiple input types including text prompts, keyframes, and reference images/videos, with various deployment options for both consumer and data-center GPUs.