For the first time, an open weight model sits at the top of a major AI video ranking. MiniMax H3, the successor to the Shanghai lab's Hailuo family, holds first place in Video Editing on Artificial Analysis, the independent benchmarking firm, ahead of every closed rival. As of Monday the claim comes with substance, because the model weights are now publicly available.
Artificial Analysis places H3 first in Video Editing with an Elo score of 1,130 built from more than 5,000 blind preference comparisons as of July 31, and the model also ranks second in Text to Video and third in Image to Video. No open weight model had ever led one of the firm's video categories before. The previous open weights leader, LTX-2.3, trails well behind, which makes H3 the strongest open video model by a wide margin from the moment its weights landed.
H3 is a 33 billion parameter model with a genuinely multimodal interface. A single prompt can combine text with up to nine reference images, three video clips and three audio clips, and the model generates clips of four to fifteen seconds at 24 frames per second with native stereo sound. Fine tuning on custom content is supported, which matters for studios that want a consistent character or style across shots.
The release is open weights, not open source, and the fine print matters. The MiniMax Community License restricts commercial use to companies with less than 20 million dollars in annual revenue, and two pieces of the hosted product stayed home. The 2K resolution module and the H3-Context-IR component are not included, so local runs through ComfyUI top out at 768p and users must handle context preparation on their own.
The timing underlines how crowded the video race has become. ByteDance shipped Seedance 2.5 the same day, generating 30 second clips with built in audio, and closed models from the biggest labs still dominate most of the video leaderboards. But text models have already shown how quickly open weights releases can reset a market once they reach the front of the pack, and H3 is the first evidence that video generation is heading down the same path.