FastH3-Omni is out with open weights and a ComfyUI template: 8-step reference video with sound
Hao AI Lab and NuvaLab released the weights and training code for FastH3-Omni, a distilled MiniMax H3 that turns images, video clips and audio references into a video with synced stereo sound in 8 steps instead of 50. ComfyUI supports it on day one with a built-in template and FP8 and NVFP4 files for smaller GPUs.
What it does
You can give it up to 9 images (3 in realtime mode), 3 video clips and 3 audio clips, 12 references at most, plus optional first and last frames or a clip to continue. It returns 5 to 15 seconds at 24 fps with 32 kHz stereo audio. The Hugging Face collection has BF16 weights, ComfyUI files in BF16, FP8 and NVFP4, and an MLX build for Macs.
Results and limits
The team reports 5.6 to 11.9 times lower latency than the 50-step model on 8 GB200 GPUs. Those are server numbers; the team's separate local test (see the FastH3 Omni local entry) shows much longer times on desktop cards.
Worth knowing
In ComfyUI, use the video_fasth3_ominiref_r2v template and download the listed diffusion model, Qwen3-VL text encoder, video VAE and audio VAE. It is under the MiniMax-H3 Community License. MiniMax H3: Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use.