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

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

What HyperFlow H3 Flash actually is

HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of the MiniMax-H3 transformer for joint video and synchronized stereo audio generation. The repository provides model weights (GGUF and INT8), showcase material, ComfyUI nodes, and setup guides.

The model is about 13.7-13.9B parameters (less than half of the ~31B MiniMax-H3) and produces 5.2-15 s clips at 24 fps (default 1344×768) with a 32 kHz stereo soundtrack.

What you can do and how to run it

You can run end-to-end text-to-video (and reference-image tasks) including text encoding, transformer sampling, and video+audio VAE decode. Sampling uses 4, 6 (default) or 8 steps and runs without classifier-free guidance (CFG).

Integrations and runtimes include ComfyUI (GGUF), Unsloth Studio (GGUF), stable-diffusion.cpp (patch required), and a Python package for the INT8 convrot transformer. The repo gives quick-start instructions, notes on required text encoder downloads, and commands for ComfyUI templates or python -m hyperflow_h3_flash.

Performance, resource notes, and who this fits

Performance examples: a 5.2 s clip at 6 steps runs in about 8.7 s on 4× B200, 36 s on 1× H200, and 92 s on 1× RTX 5090. Minimum footprints vary: Q4_K_M is 9.1 GB; Q8_0 is 14.9 GB. GGUF is recommended on PCs with 64 GB system RAM or less; INT8 convrot is targeted at data-center GPUs or 24 GB+ RTX GPUs with large host RAM.

This is aimed at people who want to run compact, few-step video+audio generation and who are comfortable with model downloads, ComfyUI or command-line setups, and managing GPU and host RAM tradeoffs.

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

Do I need special GPU or RAM to run this?

Yes - the repo documents different setups: GGUF fits better on PCs with ≤64 GB system RAM, while INT8 convrot targets data-center GPUs or 24 GB+ RTX GPUs with large host RAM; minimum model footprints and memory notes are provided.

What outputs and sampling settings are supported?

Outputs are 5.2-15 s clips at 24 fps with 32 kHz stereo audio; sampling supports 4, 6 (default), or 8 steps and runs without classifier-free guidance.

How do I integrate it into ComfyUI or other runtimes?

The repo gives quick-start steps: clone ComfyUI-GGUF and ComfyUI-HyperFlow-Flash into custom_nodes, download the nvfp4 text encoder, start ComfyUI with --fast-disk on 16/32 GB PCs, and open the provided workflows.

Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8 was shared on NexURLs by NexURLs on November 8, 2026, in Science.