MiniMax announced H3 on X on July 31, 2026, calling it an open model that breaks the boundaries between tasks and modalities. The company says H3 understands unified context across text, images, video and audio, and generates up to 15 seconds of video at 2K resolution with native stereo sound. That detail matters: most video generators still bolt audio on after the fact.
MiniMax says H3 excels at instruction following, accurate text and brand rendering inside generated frames, and video-to-video motion transfer, the ability to take motion from one clip and apply it to a different subject or scene. The company reports that early testing shows the model ready for commercial content creation across a range of use cases. It did not specify which testers, or what the testing measured.
Native stereo audio generated alongside video is harder than it sounds. A model has to keep sound sources spatially and temporally locked to what is happening on screen. Footsteps need to land on the frames where a foot hits the ground. Dialogue needs to move between left and right channels as a speaker turns, and room tone needs to shift as a camera cuts. Most systems avoid the problem by generating video first and scoring it with a separate audio model, which produces sound that is plausible but not causally tied to the pixels. If H3 genuinely couples the two generation processes, that is a real technical step past the current state of the art in short-form video generation.
MiniMax also cites placements on the Artificial Analysis leaderboard: first in Video Editing with audio, second in Text to Video with audio, and second in Image to Video without audio. Those rankings come from MiniMax’s own announcement, not from an independent audit of the leaderboard at the time of publication, and Artificial Analysis rankings shift as new models are submitted and retested.
The announcement calls H3 an open model with open weights, but it does not include a link to a weights repository, a model card, or a license. In the replies, users asked where the weights actually were and how many parameters the model has. MiniMax did not answer either question. A model described as open but not yet downloadable, with no disclosed parameter count, is not yet verifiable as open. It is a claim of intent to open it, not evidence that it has happened. That distinction matters more than usual here because parameter count and licensing terms are exactly what determine whether a team can run H3 on its own infrastructure, fine-tune it, or budget for inference costs.
Teams evaluating H3 for commercial video production should treat today’s post as a capability preview, not a release, and wait for a published weights link and license before scheduling it into a 2026 production pipeline.
MiniMax announced H3 in a post on X on July 31, 2026.
