StockDuty ← Dashboard Nov 07, 2026 06:14 PM ET
SignalTech/AI

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

Compression from 31B to 14B with 6 steps and INT8 quantization drives high relevance.

What happened: Video Rebirth released HyperFlow H3 Flash, a compressed few-step version of the MiniMax-H3 video and audio diffusion model, with weights on Hugging Face plus ComfyUI nodes and a GGUF setup guide. Key numbers: 13.7B (fl2va) or 13.9B (ref2va) parameters versus the ~31B MiniMax-H3 transformer. Output is 5.2 to 15 s at 24 fps, 1344x768 default, with 32 kHz stereo. Why it matters: The model runs on 24 GB and 32 GB consumer GPUs in PCs with 16 GB of system RAM, and NVIDIA SOL-H3 runs it in real time on 4x B-series GPUs. The 62 GB base transformer is never downloaded.
Video Rebirth released HyperFlow H3 Flash, a compressed few-step version of the MiniMax-H3 video and audio diffusion model, with weights on Hugging Face plus ComfyUI nodes and a GGUF setup guide.
The model runs on 24 GB and 32 GB consumer GPUs in PCs with 16 GB of system RAM, and NVIDIA SOL-H3 runs it in real time on 4x B-series GPUs. The 62 GB base transformer is never downloaded.
Sources