hyperflow-h3-flash: 92 seconds on RTX 5090, 14B video generation model

HyperFlow H3 Flash, a compressed version of MiniMax-H3, simultaneously generates video and spatial audio from text or images. It outputs clips ranging from 5.2 to 15 seconds at 24fps and 1344×768 resolution, with 32kHz stereo sound automatically synchronized. It supports low-step sampling of 4, 6, and 8 steps without CFG, enabling fast inference.

Reduced to a size of 13.7–13.9B, less than half of the original model's 31B parameters. With Q4_K_M quantization, it has a size of 9.1GB and runs on consumer GPUs equipped with 16GB system RAM and 24GB or 32GB VRAM. It supports various input methods, including text prompts, first/last keyframes, and up to 12 reference images.

Generating a 5.2-second clip takes 92 seconds on a single RTX 5090 GPU, 36 seconds on an H200, and is shortened to 8.7 seconds with a combination of 4 B200 GPUs. It is compatible with major open-source ecosystems such as ComfyUI, stable-diffusion.cpp, and Unsloth Studio, and also provides an INT8 convrot package for data center GPUs.

It follows the MiniMax H3 community license, requiring prior approval for use in the EU, UK, Korea, and US regions. It is suitable for developers who need to generate high-quality video and audio together in low-cost environments.