MiniMax-H3 Omni-Modal Generative Model Released on Hugging Face
MiniMax has released MiniMax-H3, a general-purpose omni-modal generative model, on Hugging Face with open weights. The model supports the unified understanding of text, image, video, and audio, and can generate 2K videos with native stereo audio. The release provides the open-source community with a highly capable model that generates synchronized video and audio in a single pass. This reduces the need for separate audio generation tools and advances the development of unified multimodal AI systems. MiniMax-H3 can generate videos up to 15 seconds long at 2K resolution, with speech and sound effects precisely timed to on-screen actions. Its task-generalization-oriented design allows it to follow complex multimodal instructions directly from the pre-training stage.
## BACKGROUND
MiniMax is an artificial intelligence company based in Shanghai, China, recognized as one of the country's prominent "AI Tigers." The company is known for developing multimodal AI models and consumer applications, such as the video-generation service Hailuo AI. Historically, generating video with matching audio required separate models for visual synthesis and sound generation, which often led to synchronization issues.