MiniMax H3 （自部署）

MiniMax H3 is a general-purpose, omni-modal video generation model developed by MiniMax. It unifies the understanding of text, images, video, and audio inputs, and is capable of generating videos up to 15 seconds in length at 2K resolution with native stereo audio. Unlike earlier systems that separated tasks such as text-to-video, image-to-video, and editing into distinct models, H3 treats them as a single generalized task, using natural language to describe how reference materials relate to the target output.