MiniMax launched its new multimodal generation model MiniMax H3, which accepts text, images, video, and audio inputs and outputs HD video with native stereo audio. It integrates multiple tasks into a single pretraining framework, boosting end-to-end throughput by 30% and reducing inference costs. Full open-source release is planned within days.