MiniMax H3 is a 2K reference-to-video model that combines reference images, videos, and audio with natural-language instructions to create coherent video scenes.