MiniMax H3 is first open model to top AI video ranking

Chinese AI startup MiniMax has released the weights for its H3 video generation model, marking a significant milestone as the first open-weights system to lead a major industry ranking. According to evaluations by Artificial Analysis, the 33-billion-parameter model ranks first in Video Editing, second in Text-to-Video, and third in Image-to-Video. This release represents a shift in the AI video landscape, which has previously been dominated by entirely closed proprietary systems.

The H3 model is highly multimodal, processing text, images, video, and audio simultaneously to generate video clips ranging from four to 15 seconds in length, complete with stereo sound. Its model card indicates that a single prompt can ingest a complex mix of reference materials, supporting up to nine reference images, three video clips, and three audio clips at once. On the same day, competitor ByteDance launched its own closed model, Seedance 2.5, which produces longer 30-second clips with integrated audio.

Despite the open release, MiniMax has kept two key components proprietary. The 2K resolution module and H3-Context-IR—a tool that structures prompts and reference materials—remain closed. Consequently, practitioners running H3 locally via ComfyUI will find their output resolution capped at 768p. Users must also manage their own context preparation manually by following MiniMax's published prompting documentation.

For AI practitioners and developers, the availability of these open weights opens up new avenues for customization. Teams can now fine-tune the model on their own custom footage, specific characters, or unique visual styles. However, commercial deployment comes with a financial threshold: MiniMax's license restricts commercial use of the model to businesses generating less than $20 million in revenue.