MiniMaxAI/MiniMax-H3 · Hugging Face

MiniMax H3 is a versatile open-source multimodal generative AI system capable of understanding and generating video, audio, text, and images, supporting up to 2K resolution video with stereo audio for durations of 4 to 15 seconds. It includes multiple model variants allowing different input modes, such as text-to-video or multi-reference inputs combining images, audio, and video. The system architecture features three main modules: H3-Context-IR for multimodal input understanding and instruction parsing, H3-Base for initial audio-video generation at 768p, and H3-Regenerate-2K for high-resolution output refinement. Safety measures include automated moderation of user content, and comprehensive APIs and apps provide easy access to the model globally.