NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's New Open Image Model Is Free to Download, and Not Licensed in the US

MiniMax's H3-Image makes 16-megapixel images in about ten seconds and ships its weights on Hugging Face. The licence quietly excludes the US, EU, UK and South Korea — including the images it produces.

2026-11-19 · 1189 words · NeuroAI
China's New Open Image Model Is Free to Download, and Not Licensed in the US

The weights are on Hugging Face. Anyone can download them.

The licence, though, has a map attached, and the map has holes in it. If you are in the United States, the European Union, the United Kingdom or South Korea, the model is published but not licensed to you — and neither are the pictures it makes.

That is the strangest part of the H3-Image release, and the part most likely to matter to developers outside China. The model itself is a serious piece of engineering. The paperwork around it is a policy statement.

Key takeaways

  • MiniMax released H3-Image, a 33.1-billion-parameter image generation model (图像生成模型), on November 10, with weights published on Hugging Face.
  • It produces native images of about 16 megapixels — 4096×4096 at 1:1, up to 6272×2688 at 21:9 — without a separate upscaling step, and accepts up to 16 reference images in a single edit.
  • The company claims a full 4K image in under 10 seconds; independent measurements so far range from about 12 seconds to under a minute depending on hardware.
  • The community licence excludes the US, EU, UK and South Korea from both the weights and the outputs, requires written permission above US$1 million in annual revenue, and forbids using outputs to improve other models.

What was actually built

H3-Image is a diffusion transformer — a DiT — with 33.1 billion parameters in a dense, single-stream design, 50 layers deep. It is derived from MiniMax's H3 video model with the audio branch removed, and it reuses H3's Qwen3-VL text encoder unchanged.

Two decisions follow.

The first is that generation and editing share one checkpoint. You can write a prompt and get an image, or hand it up to 16 reference images plus a prompt and get a recomposed scene. For anyone doing character sheets, product shots or storyboards, that is the difference between a toy and a tool, because consistency across images is where most workflows actually break.

The second is that the model was distilled down to six sampling steps, with no classifier-free guidance and therefore no negative prompt. Fewer steps means less compute per image, which is how MiniMax gets to its speed claim. The trade-off is control: you cannot dial guidance up or down, and early testers report that small text and fine details still slip.

Native resolution is the headline. Most image models generate at one to four megapixels and then upscale. H3-Image generates at about 16 megapixels directly, across aspect ratios from 1:3 to 3:1. The model card puts the useful working range at roughly 4 to 20 megapixels.

The speed claim, and what it is worth

MiniMax says a 4K image takes under 10 seconds. On a like-for-like basis, it quotes Standard mode at 1.5 to 2 times faster than mainstream alternatives and Highspeed mode at 3 to 4 times — the larger multiple belongs to the tier that produces the sub-10-second figure.

Those are the company's numbers, and they are not consistent. Standard mode is quoted at around 16 seconds in some launch coverage. Independent and community measurements have ranged from about 12 seconds on a hosted service, to 22–24 seconds through the API, to 55 seconds on a single RTX 3090 with quantised weights.

The honest summary is that "4K in ten seconds" is achievable on the right hardware and tier, not a property of the model in general.

The licence is the story

The weights are free. The licence is not open source in the sense most developers mean.

The MiniMax-H3-Image Community License defines an "Applicable Territory" as worldwide excluding the "Excluded Territories" — the European Union, the United Kingdom, the Republic of Korea and the United States of America. The grant is made "solely within the Applicable Territory". A later clause extends the restriction to outputs: you may not use, reproduce, modify, distribute or display the model or "any of their Outputs or results outside the Applicable Territory".

Read plainly, that covers a screenshot in a US blog post and a generated illustration in an EU ad, not just running the weights on a local GPU.

Three more terms matter everywhere else. Companies whose annualised revenue exceeds US$1 million, or anyone offering the model to third parties as a hosted service, need prior written authorisation. Commercial interfaces must display the model name. And outputs may not be used to improve any other AI model, with the licence explicitly treating distillation and synthetic-data training as creating derivatives. Hong Kong law governs.

MiniMax has not published a detailed rationale for the image model. For the earlier H3 video model, which carries the same four exclusions, its licence Q&A pointed to the EU AI Act, unsettled rules for AI video in the UK and South Korea, and copyright litigation in the US, and described the restriction as "not yet" rather than "not ever". Companies in those regions can apply for a separate licence.

What this means for the open-weights narrative

Chinese labs have spent two years arguing that open weights are how you win global developers. H3-Image is a reminder that "open" is now a spectrum with a customs border.

The practical effect is a split market. A developer in Singapore or Brazil can download about 135 gigabytes of weights — 68.65 GB of DiT plus a 66.7 GB text encoder — and run the model locally. A developer in Berlin or Austin cannot, unless MiniMax grants permission, and must fall back on the hosted API, which is governed by separate platform terms.

Honest limitations

I have read the licence text published with the model and MiniMax's own launch post, and cross-checked specifications against the model card as reported by several independent write-ups. I have not run the model, and I am not a lawyer; the territory analysis above is a reading of the licence, not legal advice.

There is no published technical report and no third-party benchmark for H3-Image as of this writing. MiniMax's claim of best-in-class prompt adherence among open-weight models is not backed by a public comparison against named rivals. The speed numbers come from the vendor, a hosting provider and individual testers using different hardware, precision and prompt lengths, so they are not directly comparable.

What readers can do now

  • Read the licence before you check the VRAM. The hardware question is interesting; the territorial question can end the project. For teams in the excluded regions, the realistic paths are the hosted API or a written licence request.
  • Benchmark on your own prompts. Six fixed steps, no guidance scale and no negative prompt mean the model is not tunable in the usual way. Test small text, hands and your specific product categories before committing a pipeline.
  • Treat "open weights" as a category, not a promise. Ask three questions of every release: which territories, what revenue threshold, and whether outputs are covered by the same terms as the weights.

A 16-megapixel image in ten seconds is a real engineering achievement. The more durable signal is that the licence, not the benchmark, is where the strategy now lives — and a download link is no longer the same thing as permission.

Related coverage

Cite this article

NeuroAI (脑机网), "China's New Open Image Model Is Free to Download, and Not Licensed in the US," 19 November 2026, https://neuroai.site/a/na-model-minimax-h3-image-licence

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories