MiniMax H3, the open-weights video model: what it does, and under which licence
MiniMax H3 is the video model from MiniMax, released in open weights under a community licence that excludes the European Union, the United Kingdom, South Korea and the United States, renders included. What it can do, what Axiris has measured, and who it can serve.
MiniMax H3 is a video generation model developed by MiniMax, released in open weights under the MiniMax H3 Community License. It produces a shot from a text, a start image, a start and end image, or character references, up to nine, with sound in the same pass. Its licence excludes the European Union, the United Kingdom, South Korea and the United States, and the prohibition also covers the videos produced. At Axiris, it has been the engine for shot-by-shot video since August 2026; video for everyone goes through the cloud models.
An open-weights video model
MiniMax H3 is the video generation model from MiniMax. Its weights are published: you can download it and run it yourself, which sets it apart from video models accessible only through an API. That does not make it open source: it is open weights under conditions, and the conditions matter.
What it can do
A shot is generated from a text alone, from a start image, from a start image and an end image, or from reference images. It accepts up to nine references per shot: a character, a setting, an object. Sound is written in the same gesture: an atmosphere described per shot comes out with the picture.
MiniMax publishes a writing format for its prompts: six sections, a vocabulary of camera moves, twelve moves, two amplitudes, two speeds, written into the sentence. A shot is described as a shot on a film set, not as an image.
It does not read a whole storyboard. You give it one shot at a time: that is precisely what the Axiris timeline does.
The licence: read before using
The MiniMax H3 Community License defines excluded territories: the European Union, the United Kingdom, the Republic of Korea and the United States. Outside the applicable territory, you may neither use the model nor display or distribute its outputs. The prohibition covers the videos produced, not only the weights. A "Powered by MiniMax H3" attribution is mandatory on the interface that uses it. Permission can be requested from MiniMax.
The text is what counts, not this page: the licence on the model page. On reading a model licence in general: AI model licences.
The consequence at Axiris: no production on MiniMax H3 is offered to a client established in these territories.
What Axiris has measured
Since August 2026, MiniMax H3 has been the engine for shot-by-shot video in the studio: Video, shot by shot. What we learned by measuring it:
One reference image, one body. A character sheet with several views pasted together produces silhouettes in the setting; the same views as separate images produce none, and nine references cost barely more than one.
The text on a reference gets printed. A title on a character sheet ends up baked into the image; the sheet is cropped down to the figures alone.
An area redone on the start image carries into the video. An exact element placed in the start image is animated with the scene, reflection included. An element pasted flat, which the scene does not believe in, disappears.
A character's identity holds across nine shots generated separately from a single sheet. That is judged by eye: no likeness measurement exists yet, and we do not present this result as a figure.
Its place in the offer
MiniMax H3 serves the studio's shot-by-shot route, for those who can use it. Video for everyone, from the same interface, goes through the cloud video models, which take a whole storyboard and render the film: Seedance Pro, Cloud of your choice, with Nova. The route is chosen according to the project and the client's country.
The other engines the studio uses, with their date: Models of the moment.