MiniMax H3

MiniMax H3 is a multimodal video generation model developed by MiniMax, released in July 2026. It is designed to process and generate synchronized audiovisual content, supporting native 2K resolution video and stereo audio in a single generation pass. The model is built on an "omni-modal" framework that allows it to ingest a unified context of text, images, video, and audio simultaneously, rather than processing these inputs through separate, sequential modules.

The model's architecture incorporates the H3-Omni Transformer, H3-VAE, and Contextual Omni Representation technologies. This allows the system to understand complex relationships between diverse reference materials. For example, a user can provide an image to define a character's appearance, a video clip to define motion or camera behavior, and an audio file to guide voice or sound effects, all of which the model integrates into a cohesive 5-to-15 second video clip.

Key capabilities of MiniMax H3 include Omni Reference conditioning, which supports up to 12 mixed files as input (including up to 9 images and 3 video/audio clips). It excels at text-to-video, first-and-last-frame image-to-video, and video-to-video motion transfer. Additionally, the model supports multi-shot generation, enabling it to produce sequences with distinct cuts, transitions, and synced sound from a single descriptive script or brief.

To achieve the best results, the official guidance suggests structuring prompts to describe the action, camera movement, and sound environment in detail. When using multiple references, clearly defining the role of each input—such as which image defines the subject and which video defines the movement—improves the model's instruction-following accuracy. The model natively outputs video at 24 frames per second and supports various aspect ratios including 16:9, 9:16, and 1:1.