MiniMax released H3 today, a multimodal video generation model ranked first in video editing by independent benchmarking firm Artificial Analysis, and announced plans to release its weights publicly, which would make it the most capable open-weight video model ever shipped. The release lands while the Hailuo AI platform faces an active copyright lawsuit from Disney, Universal, and Warner Bros. Discovery that survived a federal judge's motion-to-dismiss ruling in May 2026. H3 generates video clips of 4 to 15 seconds at native 2K resolution (2,560×1,440 pixels) and 24 frames per second, with stereo audio produced in the same inference pass as the video. The model accepts up to nine reference images, three video clips, and three audio tracks in a single generation request, which MiniMax calls the "omni-reference" system. It also supports instruction-based editing, allowing creators to specify changes to part of an existing clip while leaving the rest intact. The engineering mechanism behind H3's price advantage is a rebuilt video tokenizer called H3-VAE, which compresses raw video frames into a latent representation, determining compute needed per frame and generation cost. Prior Hailuo models used a standard tokenizer biased toward low-level pixel detail.