MiniMax releases open-weight H3 video model, day-zero support lands in ComfyUI for local use
AI video model releases
MiniMax releases open-weight H3 video model, day-zero support lands in ComfyUI for local use
MiniMax released the weights for its H3 video model on August 2, and ComfyUI added built-in support the next morning, letting users run it locally instead of through a paid cloud service.
ComfyUI engineers cut the smallest variant's memory footprint from 123.6 GB to 42.5 GB using parameter pruning, 8-bit compression, and dynamic offloading to system RAM, making it runnable on an Nvidia RTX 3060.
H3 combines five tasks in one model: text-to-video, image animation, first/last-frame control, reference transfer of subjects, motion, or voices, and editing existing clips without regenerating from scratch.
A single task can take up to nine reference images, three video clips, and three audio clips, producing 4 to 15 second clips at 24 fps with stereo audio.
The release follows a pattern set by DeepSeek, Alibaba Qwen, and Moonshot, where Chinese labs use open weights and low prices to compete against US firms that mostly keep top models behind closed cloud services.
cho.sh