MiniMax's Open-Weight H3-Image Makes 4K Pictures in Seconds, Under a Licence That Shuts Out the US, EU, UK and Korea
MiniMax released H3-Image on November 11, an open-weight text-to-image and image-editing model built on its H3 family, with a 33.1 billion-parameter diffusion transformer distilled to six sampling steps. MiniMax says it generates a 4096x4096 image in under 10 seconds and accepts up to 16 reference images; those are its own claims. The weights are on Hugging Face under a community licence whose territory excludes the European Union, the United Kingdom, South Korea and the United States, and which requires written authorization above $1 million in annual revenue or for hosted services.
MiniMax's new image model is open weights for most of the world, but not for anyone in the US, the EU, Britain or South Korea. H3-Image, released on November 11, is a text-to-image and image-editing model whose weights are on Hugging Face. Its licence defines the permitted territory as "worldwide, excluding" the European Union, the United Kingdom, the Republic of Korea and the United States.
What shipped#
H3-Image is built on MiniMax's H3 family. According to its model card, it is a 33.1 billion-parameter dense diffusion transformer with the H3 architecture minus the audio layers, uses the MiniMax-H3 text encoder unchanged, and is distilled to six sampling steps with no classifier-free guidance. One checkpoint handles both generation and editing, and it accepts up to 16 reference images.
MiniMax says it generates a 4096 x 4096 image in under 10 seconds, at "up to 5x the speed of mainstream alternatives," and works best at about 16 megapixels without an upscaler. It also claims "state-of-the-art prompt adherence among open-weight models." These are MiniMax's own figures; we did not find an independent benchmark of H3-Image. The model is also sold through the MiniMax API, MiniMax Code and Hailuo AI, which MiniMax says is priced at 50 to 70 percent of comparable offerings, with 50 percent off 4K requests in launch week.
Running it locally is not light: the transformer weights alone are about 66 GB in BF16, and the text encoder, loaded from the separate MiniMax-H3 repository, is another 66.7 GB, the model card says.
The licence is the story#
The MiniMax-H3-Image Community License Agreement grants rights "solely within the Applicable Territory" and bars use of the model or its outputs outside it. Even inside that territory, users whose annualized revenue, with affiliates, exceeds $1 million, or who offer hosted services to third parties, need separate written authorization from MiniMax. Commercial products must display "MiniMax-H3-Image" in their interface, and outputs may not be used to improve other AI models. MiniMax says it will keep evaluating the excluded markets and invites people there to apply for a licence.
That is tighter than MiniMax's own recent text release: M3.1 Flash's community licence set its written-authorization threshold at $20 million in yearly revenue, twenty times higher.
Limits#
- Speed, quality and prompt-adherence claims are MiniMax's; we found no independent leaderboard result at publication and did not run the model.
- We read the licence text on Hugging Face; MiniMax has not explained publicly why those four territories are excluded.
- All sources here are MiniMax's own pages.