AI
HyperFlow H3 Flash Compresses 31B Audio-Video Diffusion Model to About 14B
Video Rebirth released HyperFlow H3 Flash, a compressed fewer-step version of MiniMax-H3 that can simultaneously generate video and synchronized stereo audio. The model has approximately 13.7B to 13.9B parameters, less than half the original roughly 31B transformer, uses 6-step sampling by default, and requires no CFG.
HyperFlow H3 Flash can generate 5.2- to 15-second, 24fps videos at a default resolution of 1344×768, accompanied by a 32kHz stereo audio track. Sampling supports 4, 6, and 8 steps, with 6 steps by default, and does not use classifier-free guidance. The project provides model weights, a ComfyUI node, a GGUF setup guide, and an 8-step LoRA for the full MiniMax-H3.
In terms of speed, when generating a 5.2-second clip with 6-step sampling, 4 B200 GPUs require 8.7 seconds, a single H200 requires 36 seconds, and a single RTX 5090 requires 92 seconds. Inference acceleration partner NVIDIA SOL-H3 says the model can run in real time on 4 B-series Blackwell GPUs. The model has 13.7B or 13.9B parameters depending on the fl2va or ref2va version; Q4_K_M requires a minimum of 9.1GB and Q8_0 requires 14.9GB. It can run on personal computers with 16GB of system memory and a 24GB or 32GB GPU; measurements for a 16GB GPU have not yet been completed.
The project recommends GGUF for personal computers with no more than 64GB of system memory; INT8 convrot is recommended for data-center GPUs or RTX cards above 24GB with at least 69GiB of system memory. Software requirements include Python 3.10 or later, PyTorch 2.6 or later with CUDA and Triton, an Ampere or newer NVIDIA GPU, and diffusers 0.40 or later. The first run downloads about 90GB, or about 68GB when using fp8 text encoder precision, including the text encoder, VAE, and an approximately 16GB HyperFlow INT8 transformer for the selected task.