MiniMax posts H3-Image for a 4K frame in under 10 seconds, and the license leaves out the U.S. and Europe
The research blog is dated 11 November. The Hugging Face license is dated 10 November, names Nanonoble Pte. Ltd. as the licensor, and requires written permission once annual revenue passes $1 million or the weights are hosted for someone else.
Feature2 minIn-depth
James Okonkwo2 min read
- MiniMax
- Open weights
- Image models
- Qwen

Photo: MiniMaxSource page
MiniMax on Wednesday published MiniMax H3-Image, an open-weight image model it says is built on H3. The research post is dated 11 November 2026. It supports text-to-image generation, generation from reference images, and editing, and the company says it can produce a 4K frame, 4096 by 4096, in under 10 seconds, using up to 16 reference images. The weights are on Hugging Face. The same post says the model is also on the MiniMax API, in MiniMax Code, and in Hailuo AI.
The speed and price lines are the company’s, and they do not name the models they are measured against. H3-Image, the post says, generates a full-resolution 4K image at up to five times the speed of mainstream alternatives. Standard mode is 1.5 to 2 times as fast as those alternatives, and a high-speed mode is 3 to 4 times as fast. Pricing is 50 to 70 percent of comparable mainstream offerings. For launch week, generation at 4K and above is half the standard price. The post also says prompt adherence is state of the art among open-weight models. It does not attach a table.
A license dated the day before the post
The community license on the Hugging Face repository is dated 10 November 2026, a day before the blog post. It defines MiniMax as Nanonoble Pte. Ltd. The grant covers the world except the European Union, the United Kingdom, South Korea, and the United States. Someone in those places is told to contact the company for a separate license. A commercial clause requires prior written authorization if the user and its affiliates have annualized revenue above $1 million, or if the user offers a hosted service to third parties. Outputs may not be used to improve any other AI model, other than H3-Image or a derivative of it. The prompt-enhancer model bundled with the release is described as derived from Qwen3.8-27B, under the Apache 2.0 license. Disputes are under Hong Kong law.
The acceptable-use exhibit, revised the same day as the license, is a long list. It bars military use, impersonation without consent, and putting machine-generated images, code, or posts into a public setting without a clear disclosure that they were machine-generated. It also bars using the model to improve a different model by the route already closed in the body of the license. The Wednesday post says the weights are open so that developers can customize them. The license is the document that says where, and for whom, that is allowed.


