MiniMax H3 model generates 15-second 2K video

MiniMax released its new multimodal model H3 on July 31, which can understand text, images, video, and sound, and output native dual-channel audio-video content, supporting up to 15 seconds at 2K resolution.

Early testing shows H3 excels in instruction following, text and brand presentation, and video-to-video motion transfer, suitable for advertising, e-commerce, brand design, UI/UX, and gaming. It adds a captioning capability with a custom pipeline, using about 100K tokens for most inputs and outputting around 4K tokens.

The 2K video output doesn't rely on traditional upscaling but uses the base model to regenerate low-res results, preserving details that conventional methods miss. The company plans to open the model weights in the coming days.