MiniMax H3, a video generation AI, has been unveiled and is scheduled to be released as an open model, boasting the world's second-best performance.
MiniMax, a Chinese AI development company, has announced its video generation AI, ' MiniMax H3 .' MiniMax H3 can generate videos with sound up to 15 seconds in length and supports text, video, image, and audio input. The model itself is also scheduled to be released for free in the near future.
MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities - MiniMax Research | MiniMax
MiniMax H3 supports text, video, image, and audio input, allowing you to perform operations such as 'inputting images to specify characters' or 'inputting singing voices to create lip-sync videos.' You can also input multiple media simultaneously and generate a single video. Video resolution is 768p or 2K, and you can generate videos up to 15 seconds long.
The following is an example of generating a MiniMax H3, as shown on the official MiniMax website.
There are also videos that look like game screens.
Video generation example #2 using the AI 'MiniMax H3' - YouTube
Test results from a third-party organization, Artificial Analysis , have already been published. The following is a ranking of the quality of the task of 'generating videos with audio from text' using the method of 'generating videos from the same prompt with the AI model's name hidden to determine superiority.' MiniMax H3 ranked second, following Gemini Omni Flash. It surpasses Seedance 2.0, which is known for its high quality.
For the task of 'generating a video with sound from an image,' the top performers were Seedance 2.0 (1st place), Gemini Omni Flash (2nd place), and MiniMax H3 (3rd place).
MiniMax H3 is available via API at the time of writing, and the model will be released as open weights soon. The API documentation can be found at the following link.
Create Video Generation Task - MiniMax API Docs
https://platform.minimax.io/docs/api-reference/video-generation-v2-create
Furthermore, third-party services such as ComfyUI can also generate videos using the MiniMax H3 API.
MiniMax H3 is now available in ComfyUI via Partner Nodes.
— ComfyUI (@ComfyUI) July 31, 2026
→ Multimodal I/O: T2V, First/Last Frame, Omni Reference
→ Native stereo audio on every clip
→Up to 2K resolution, 5-15 seconds at 24 FPS
→ Edit characters, scenes, dialogue, and voice in place
API access is live today. Native… pic.twitter.com/lRrgrIEvFN
◆Added on August 3, 2026 at 12:01 PM:
The model has been released on Hugging Face and ModelScope.
MiniMaxAI/MiniMax-H3 · Hugging Face
https://huggingface.co/MiniMaxAI/MiniMax-H3
MiniMax-H3 · Model Garden
https://modelscope.cn/models/MiniMax/MiniMax-H3
It is also available in the Hugging Face repository managed by the ComfyUI development team. The file sizes are 21GB for the int8_convrot version of the diffusion model and 27.1GB for the int8_convrot version of the text encoder. The VAEs are 5.21GB for video and 605MB for audio.
Comfy-Org/MiniMax-H3 · Hugging Face
https://huggingface.co/Comfy-Org/MiniMax-H3
Related Posts:
