MiniMax H3 ranks #1 on Artificial Analysis, but open weights miss 2K upscaling
minimax released the 33b parameter h3 video model under a restrictive open license, ranking first in video editing and second in text-to-video on artificial analysis. this challenges closed-source dominance in video generation.
minimax released the h3 video model, a 33-billion-parameter system capable of processing text, images, video, and audio inputs to generate clips up to 15 seconds long with stereo sound. on the artificial analysis leaderboard, h3 ranked first for video editing, second for text-to-video, and third for image-to-video, outperforming many closed rivals. the model weights are open for local use and fine-tuning, but the most critical component, the 2k resolution upscaling module, remains proprietary, and commercial use is restricted to companies earning under $20 million. this release signals a shift where open-source video models can match or exceed proprietary options on specific tasks, potentially accelerating adoption for bespoke commercial applications.