MiniMax H3

MiniMax H3 is a general-purpose omnimodal generative model released by MiniMax in July 2026, designed to move beyond single-modality tools toward a system that understands complete creative intent across text, image, video, and audio . It generates video up to 15 seconds at 2K resolution with native dual-channel audio, and handles complex tasks like video-to-video motion transfer, in-context editing, and multimodal reference . Rather than treating generation and editing as separate pipelines, H3 processes mixed-modal context to produce or modify content in one pass .

The model was open-sourced shortly after release and saw over 24 million downloads within three weeks, with more than 300 public derivative models built on it—making it the most-downloaded model globally in 2026 by that measure . MiniMax has positioned H3 as particularly strong at adjusting existing video rather than generating from scratch, which some observers described as an "AR-like" approach to reality augmentation rather than pure virtual creation . The company has since pushed H3 into real-time applications including 24-hour AI livestreams and Twitch broadcasts through a faster variant, H3 Max, which can generate 5 seconds of 768p audiovisual content in under 3 seconds .