31B video-audio diffusion model compressed to 14B
Video Rebirth released HyperFlow H3 Flash, compressing the roughly 31B MiniMax-H3 to 13.7B/13.9B parameters with 4/6/8-step, CFG-free sampling that generates video plus 32 kHz stereo audio. A 5.2-second clip takes about 8.7 s on 4× B200 and 92 s on one RTX 5090; Q4_K_M is just 9.1 GB.