MiniMax H3 video ranking leads editing, but open-model claim needs qualification

MiniMax H3 took the top spot in Artificial Analysis’ video-editing leaderboard on July 31, while placing second in text-to-video and third in image-to-video. The MiniMax H3 video ranking gives the Shanghai company a credible category win in a market led by closed systems, but it does not establish the broader claim that H3 was the first open model ever to lead an AI video ranking.

Artificial Analysis said its evaluation used H3’s 2K tier and ranked the model first for instruction-based video editing. Its July 31 post also said MiniMax planned to release the model weights under the MiniMax Community License. That timing matters: the leaderboard result preceded the reported weight release.

By August 3, Global Times reported, citing a MiniMax statement, that the company had released H3’s weights. The evidence available does not include a complete historical record from Artificial Analysis or another ranking provider that could verify a first-ever claim. Nor was H3 the leader across the other two tested video categories.

What did MiniMax H3 rank first for?

H3 ranked No. 1 in Artificial Analysis’ Video Editing leaderboard. The benchmark placed it No. 2 in Text-to-Video and No. 3 in Image-to-Video. South China Morning Post separately reported that H3 trailed Google’s Gemini Omni Flash for text-to-video and ranked behind both ByteDance’s Seedance 2.0 and Gemini Omni Flash for image-to-video.

For operators, that is a narrower result than a claim of overall video-model leadership. It suggests H3’s strongest measured position was editing existing video through instructions and supplied reference material, rather than generating video from text or images alone.

What is available, and what remains restricted?

Reuters reported that MiniMax launched H3 on July 31 as a model that can process text, images, video and audio. MiniMax said it can generate clips of up to 15 seconds at 2K resolution with native stereo sound, as well as edit content and transfer motion between videos. Those are company claims, not independently verified performance results.

Artificial Analysis said H3 was available through the Hailuo AI app and MiniMax’s API at the time of its ranking post. MiniMax has positioned the model for advertising, e-commerce, product design and games, and claims 2K generation costs less than one-third of mainstream alternatives. Reuters reported those price and use-case assertions as MiniMax’s own.

The release should not be treated as an unrestricted open-source distribution. Artificial Analysis said the planned MiniMax Community License allows commercial use only by organizations with less than $20 million in revenue and requires prominent attribution. The Decoder reported that the 2K-resolution module and H3-Context-IR, a component used to turn prompts and reference assets into structured intermediate context, were not included in the released package. It said local use in ComfyUI is limited to 768p.

That leaves MiniMax with a differentiated distribution pitch: developers below the revenue threshold can download and customize substantial model weights, while the highest-resolution workflow and part of the context-processing stack remain controlled by the company. Whether that arrangement changes adoption will depend on implementation, licensing compliance and results outside benchmark evaluations.

This story draws on original reporting from The Decoder.