MiniMax released the weights for its H3 video model, and an open model now tops a video ranking for the first time
MiniMax published the weights for H3, a 33 billion parameter video model that handles text, images, video and audio together. Artificial Analysis ranks it first in video editing. Two components stayed closed, and the licence has a revenue cap.
Chinese lab MiniMax has published the weights for H3, its newest video model, and for the first time a freely downloadable model sits at the top of a video generation ranking. The benchmarking site Artificial Analysis places H3 first in video editing, second in text to video and third in image to video. Until now, the leaders in every one of those categories were closed, paid products.
H3 has 33 billion parameters, which is small by current standards and well within reach of a single well equipped workstation. Parameters are the adjustable numbers a model learns during training, and they are a rough proxy for size rather than quality. The model takes text, images, video and audio as input in one pass, and produces clips of four to fifteen seconds with stereo sound. According to the model card on Hugging Face, a single prompt can carry up to nine reference images, three video clips and three audio clips, which is what makes the editing score plausible: you can hand it an existing shot and tell it what to change.
Two caveats are worth knowing before you get excited. The 2K resolution module is not part of the release, and neither is H3-Context-IR, the piece that turns your prompt and reference material into a structured internal format the model actually reads. Run H3 yourself in ComfyUI, a popular node based interface for open image and video models, and you top out at 768p while doing the context preparation by hand using MiniMax’s published prompting guides. The licence adds one more limit: commercial use is only permitted for companies below 20 million dollars in revenue.
What is behind this. Open weights means the trained model file is downloadable and runnable on your own hardware, as opposed to an API where you send a prompt to somebody else’s server. Video has been the last stronghold of closed models, because generating moving images is expensive to train and expensive to serve, and because the copyright exposure is uncomfortable. What MiniMax has done here is a familiar pattern with a twist: release enough to win the ranking and the developer goodwill, hold back the pieces that make the hosted product convenient. The resolution module and the context translator are exactly the parts that turn a strong model into a smooth product. ByteDance released its closed Seedance 2.5 the same day, with 30 second clips and built in audio, which tells you the commercial race is not slowing down.
What this means for you: if you have been curious about AI video but did not want to hand footage to a cloud service, this is the first release where running it locally is a real option rather than a compromise on quality. Expect to spend an afternoon on setup and to live with 768p. For everyone else, the practical effect arrives later and indirectly: open weights push the price of hosted video tools down, and they let smaller European providers build offerings without licensing a frontier model. If you make video professionally, the revenue cap in the licence is the line to read twice.
Sources
Karpathy turned one paragraph of Tolkien into a 3D scene for ten dollars, and called it a vibe check
Andrej Karpathy gave Claude Opus 5 the opening of The Lord of the Rings and got back 5,500 lines of working 3D browser code in two hours. He is careful to say this is not a benchmark, which is the most useful part of the story.
