MiniMax H3: open-weight video with native audio, geo-locked license
Video with 32kHz stereo in one pass, now runnable in ComfyUI via a 4-step Turbo LoRA — but the license excludes the US, EU, UK, and 2K stays hosted.
Copy markdown
Picture and sound in one pass
H3 generates up to 768p, 15-second clips with 32 kHz stereo audio produced in the same pass — no separate audio model or lip-sync step. It takes text, image, video, or audio as input, with first/last-frame control and motion transfer built in.
It runs on a consumer GPU now
ComfyUI's int8-quantized, pruned repackage cuts the memory footprint 66% — 123.6 GB down to 42.5 GB — and the team ran it on an RTX 3060 with offloading, though real use wants 64 GB+ of system RAM. Needs ComfyUI 0.30.0+; weights are at Comfy-Org/MiniMax-H3.
A Turbo LoRA drops it to 4 steps
A community Turbo LoRA cuts sampling from ~20 steps to 4 — roughly 5x faster — fast enough to prototype an audio-video pipeline in a weekend. Expect softness at 4 steps; 6-8 is the current sharpness sweet spot.
Read the license before you build
The H3 Community License excludes the US, EU, UK, and South Korea, and commercial use above $20M revenue needs separate sign-off. Only the base weights ship: the Context-IR preprocessor and 2K upscaler stay hosted APIs, so local runs cap at 768p.