MiniMax H3 tops Video Arena as No.1 open video model
Video Arena placed MiniMax-H3 first among open models in both text-to-video and image-to-video, some 280 Elo points clear of the next open weights entry. The weights are now public, but the licence restricts open-weight use in the US, EU, UK and South Korea.
The result
Video Arena published updated leaderboards late on August 3 showing MiniMax-H3 ranked first among open models in both of its main categories, text-to-video and image-to-video. The margin is the notable part: the arena put H3 roughly 280 points ahead of the next-best open weights model on its Elo scale, a gap far larger than the increments that usually separate consecutive releases.
Against closed models the picture is tighter but still striking. In the image-to-video arena H3 scored 1,455 points, three points behind Muse Video at 1,458, leaving it effectively tied for third place overall. MiniMax confirmed the placement, noting the model now leads open video generation on both Video Arena and Artificial Analysis. Those are separate boards run on separate vote pools, and they do not agree in detail: on Artificial Analysis, H3 sits first in video editing while placing second in text-to-video and third in image-to-video against all comers.
Video Arena scores come from blind pairwise human votes — users compare two clips generated from the same prompt without knowing which model produced each — so the numbers measure preference rather than any fixed capability metric.
Context
H3 is MiniMax's omni-modal successor to the Hailuo line, accepting text, images, video and audio as input and returning clips with a native audio track. The weights are now published on Hugging Face under the MiniMax Community Licence — a release of roughly 42.5GB split across two task-specific checkpoints — and third-party creative tooling picked it up quickly.
The licence is not unrestricted. MiniMax currently limits open-weight availability in the European Union, the United Kingdom, South Korea and the United States, citing the regulatory picture around likeness generation, copyright and content safety, and directs organisations in those territories to an application process for formal authorisation. The hosted API remains globally available. Reporting on the release also indicates that some components, including the 2K resolution module, were held back from the public checkpoints. MiniMax says it will revisit the territorial scope as regulation settles.
The arena result is the first time an openly released video model has led human-preference rankings rather than trailing the closed leaders by a visible margin.
Why it matters
Video generation has been the most stubbornly closed corner of generative AI. Compute costs, training data exposure and the commercial value of a good video model kept the frontier proprietary long after text and image models opened up. A 280-point open-weights lead that also sits within three points of the closed top three collapses that gap in a single release.
The licence caveat is what decides who can act on it. Self-hosting is open on paper, but is currently off-limits without authorisation in precisely the markets where most Western studios and tool builders operate, leaving them on the hosted API for now. For everyone else — and for anyone who obtains territorial clearance — the calculation for paying closed-model rates on video work has just changed.