On August 3, the AI company MiniMax officially announced the open source release of its next-generation general video model MiniMax H3. As a full-modal generation system, this model breaks through the traditional single-task boundaries, and can deeply integrate and understand multimodal contexts composed of text, images, videos, and audio, demonstrating strong task generalization capabilities.

In terms of generation specifications, H3 supports output durations from 4 to 15 seconds, with a frame rate of 24 FPS and 32 kHz stereo sound. Users can choose from various common aspect ratios such as 21:9, 16:9, 4:3, and 1:1, and can also perform basic creation by default setting the shorter side to 768 pixels. With the dedicated H3-Regenerate-2K solution, the model can even output stunning images at up to 2K resolution. In terms of multilingual interaction, it can stably support 11 major languages, including Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Spanish, Portuguese, and Russian.

To adapt to diverse creative scenarios, H3 introduces flexible model versions and rich input specifications. The H3-Base-FL2VA first-last frame mode supports inputting 0 to 2 images, and can freely and flexibly handle different tasks such as text-to-video, first-frame-to-video, last-frame-to-video, and first-last frame collaboration. The H3-Base-Ref2VA full-modal reference mode, on the other hand, supports mixed input of up to 9 images, 3 video clips (total duration not exceeding 15 seconds), and 3 audio segments, with a maximum of 12 files in total, providing professional creators with unprecedented fine control.

In the underlying system architecture, the complete MiniMax H3 includes three core modules. First is the H3-Context-IR (context intermediate representation), which can deeply analyze complex multi-instructions and convert them into structures that are easy for the model to understand, serving as a key link to ensure high-quality output. Then comes the H3-Base module, responsible for generating base audio and video at 768p resolution. Finally, there is the H3-Regenerate-2K module, which achieves 2K super-resolution reconstruction by feeding back the initial results and original context. This combination preserves vivid details while significantly improving visual fidelity.

Currently, the model has been officially released under the MiniMax H3 Community License. Developers and tech enthusiasts can obtain the complete open-source code and resources through Hugging Face. Industry experts generally believe that this open-source release will further reduce the application barriers of multimodal video generation, driving film production, visual design, and AI interaction to new heights.