Open Weights, Closed Borders: MiniMax Ships Image Model With Four Markets Barred

MiniMax has released the weights of its 33.1-billion-parameter H3-Image model, but its community licence excludes users in the United States, European Union, United Kingdom and South Korea by default. ^[1][3]

The model generates pictures from written prompts and edits images using references. MiniMax says it can produce images of about 16 megapixels without an upscaler, a separate tool for increasing resolution, and can work with up to 16 reference images. The files are published in the company's official Hugging Face repository, with release materials also on GitHub. ^[1][2]

The release gives developers access to the model's learned numerical settings, known as weights. Those files let them work with the model beyond a company-run interface. But access is only half the transaction: MiniMax's custom licence controls who may use, modify, distribute or host it, and reaches the images it produces. ^[1][3]

Businesses whose annualized revenue, including affiliates, exceeds US$1 million need prior written authorization. So does anyone offering the model as a hosted service to third parties, regardless of whether they cross that revenue threshold. ^[3]

MiniMax describes H3-Image as having the "strongest instruction following among open-source models". That is a company claim. No benchmark score accompanies it in the launch materials, and independent performance has not been established.

Practical generation times and hardware requirements beyond the deployment examples also remain unresolved, as does how MiniMax will apply its territorial rules to images distributed across borders. ^[1][4]

The border follows the picture

The H3-Image Community License Agreement, dated November 10, 2026, defines its permitted territory as worldwide except the four excluded markets. People in those places may contact MiniMax to seek a separate licence. The published agreement does not itself authorize them to use the release. ^[3]

The territorial restriction extends well beyond downloading model files. Use, reproduction, modification, distribution or display of the model materials or their outputs outside the permitted territory is not authorized under the agreement. ^[3]

That means a developer considering an image for a website has a licensing question about the picture as well as the software that made it. Running the model in a permitted country does not, under this agreement, grant unrestricted rights to display its output elsewhere. The wording gives businesses a reason to check their intended audience and distribution before deployment. ^[3]

Commercial products and services using H3-Image must also display the model's name. And the licence bars using either the model or its outputs to improve another AI model, except H3-Image itself or its derivatives. A company hoping to use generated pictures as training material for a different model therefore faces a separate restriction. ^[3]

What the download contains

H3-Image's 33.1 billion parameters are the learned numerical settings that shape its output. They describe the size of its image-generation transformer, the component that produces the picture, rather than the storage needed for the complete working setup. ^[1][2]

MiniMax lists the image model at 68.65 GB. It also directs users to download a separate H3 text encoder, listed at 66.7 GB, which turns written prompts into a form the image model can process.

Together, the two listed components amount to 135.35 GB. That is file storage, not a statement of the graphics memory required to run them. ^[1][2]

The same checkpoint, the saved set of model weights, supports both making an image from text and editing with reference images. Users do not need to switch to a separately described editing model for those tasks, according to MiniMax's technical description. ^[1][2]

MiniMax says H3-Image reaches its output in six sampling steps. Each step is a successive pass that refines the picture. Reducing the number of passes is a speed-related design choice, but six steps is not a stopwatch reading: it does not tell a buyer how many seconds a job will take on their hardware. ^[1][2]

The roughly 16-megapixel output claim applies across aspect ratios, according to MiniMax.

Resolution measures the number of pixels in an image. It does not establish whether the model follows a complicated instruction, preserves details during editing or produces a convincing result. Those are the performance questions an independent comparison would need to test. ^[1][2]

Open files, conditional rights

H3-Image makes the difference between an open-weight release, where the trained model files are downloadable, and unrestricted permission unusually visible. The files are public, but the licence is a MiniMax Community License, not a standard permissive licence such as MIT or Apache 2.0. Its grant of rights comes with territorial, commercial, distribution and training conditions. ^[1][3]

For a prospective user, the useful test is concrete: can they obtain the weights, run the model where they operate, alter it, redistribute it, sell access and use its output for their intended purpose? H3-Image answers the first question with a public repository. Its licence puts conditions on the rest. ^[1][3]

The release is real. So are the limits. Developers assessing it now have two separate jobs: test whether the model delivers the image quality and speed they need, and establish whether MiniMax permits the intended deployment. Users in the excluded territories have a stated next step before proceeding: seek a separate licence from the company. ^[1][3]