MiniMax launches H3-Image with restricted license

What did MiniMax release on Nov. 10?

MiniMax released H3-Image, an open-weight image model for text-to-image generation and image editing, according to its official repository and release notes dated Nov. 10, 2026. The company says the model uses one checkpoint to handle both generation and multi-image editing, and that it is positioned as a leading open-weight image model.

The release also ties the model to a public Hugging Face distribution and to MiniMax’s own product stack, with availability through downloadable weights, the MiniMax API, MiniMax Code and Hailuo AI. MiniMax’s repository says H3-Image is designed for 16-megapixel image generation and editing, which matters for teams building consumer tools, creative workflows, and product surfaces that need high-resolution output without a separate upscaling step.

How does H3-Image work?

MiniMax says H3-Image is built around a 33.1-billion-parameter diffusion transformer and can generate about 16 MP directly in six sampling steps. The repository lists common output sizes such as 4096x4096, 4736x3552, 5024x3360, 5472x3072 and 6272x2688, while also saying custom sizes can range from 655,360 to 20,000,000 pixels.

The model supports up to 16 reference images per request, with each file allowed up to 64 MiB and side lengths between 128 and 8192 pixels. MiniMax’s API documentation says users can control output dimensions by width and height or by area tier and aspect ratio, and that the API supports streaming output.

MiniMax also says the API can expand short prompts into detailed English descriptions before generation, while local servers use the prompt as given. The repository says the downloadable model is 68.65 GB, the optional prompt expander is 17.9 GB, and the Hugging Face repository totals about 204 GB when including the original checkpoint and diffusers format.

What can founders build on top of it?

For builders, the main appeal is the combination of generation and editing in a single model, plus local-serving support. MiniMax says H3-Image can be served locally through SGLang, vLLM-Omni, diffusers and ComfyUI, while also being accessible through its hosted API. The repository says SGLang and vLLM-Omni need about 135 GB including the MiniMax-H3 text encoder, which signals that self-hosting is possible but still infrastructure-heavy.

MiniMax’s documentation also points to a prompt ceiling of up to 32,000 characters, with content beyond 4,096 tokens truncated. That gives product teams a clearer sense of how much instruction detail the model can take, especially for complex image-editing tasks or multi-step creative generation.

The company says the model generally performs best at around 16 MP and works from about 4 MP to 20 MP. For startups building image tools, that suggests a narrower operating range than a generic vision product, but one tuned for a specific high-resolution use case.

What does the community license allow and restrict?

MiniMax is distributing H3-Image under the MiniMax-H3-Image Community License Agreement, not a standard open-source license. The official license says the release date and license date are Nov. 10, 2026, and identifies MiniMax’s contracting entity as Nanonoble Pte.

The license grants rights only within an applicable territory and defines excluded territories as the European Union, United Kingdom, Republic of Korea and United States of America. Available reporting says users in those territories may seek a separate license from MiniMax.

The agreement also requires prior written authorization for hosted services and for annualized revenue above $1 million, and says commercial products or services using the model must prominently display "MiniMax-H3-Image" in the user interface. It further bars use of the model or its outputs to improve any other AI model other than H3-Image or its derivatives, while also stating that MiniMax claims no rights over users’ outputs.

That mix of open weights and licensing constraints makes the release more usable than a closed API-only product, but materially less permissive than a standard open-source distribution. For founders, the practical question is not just whether the weights are available, but whether their target market, revenue model and deployment plan fit the license terms.

What comes next for MiniMax users?

MiniMax has not only listed H3-Image in release notes but also made it available through its API documentation, suggesting the company wants the model to be used both by developers who self-host and by teams that prefer a managed API. The public repository links the model release to Hugging Face and says the weights are available under the community license.

For now, the most important follow-on issue is compliance: whether a given product falls inside an excluded territory, crosses the $1 million annualized revenue threshold, or qualifies as a hosted service to third parties. Those distinctions will determine whether a user can rely on the published license or needs separate permission from MiniMax.

If you want a related benchmark for how AI infrastructure companies are positioning around productized models, see E2B says ARR topped $100M, headline says $120M.