MiniMax's free image AI isn't licensed for the US or Europe, and firms earning over $1 million need permission

MiniMax, the Hong Kong-listed Chinese AI company, released the files for its new image generator on November 10. Anyone can download them for free from Hugging Face, the main public library for AI models. Not everyone may use them, though.

The license bars use in the United States, the European Union, Britain and South Korea. Everywhere else, a business whose annual revenue tops $1 million needs MiniMax's written permission. So does any company that wants to run the model as a service for others.

That revenue limit is the real news. MiniMax's H3 video model, released in August with the same territory exclusions, required permission only once products built on it earned more than $20 million a year. The image model's $1 million limit is one-twentieth of that, and it counts the revenue of the whole company and its affiliates. The video license also did not require permission to run the model as a service.

For buyers, the license points to a clear strategy: use free downloads to win attention, then charge almost any company of real size.

What MiniMax released

The model, MiniMax-H3-Image, creates pictures from text and edits existing images. It can take up to 16 reference images in one request. It produces images of about 16 megapixels directly, such as a square of 4,096 by 4,096 pixels, with no separate step to enlarge them.

MiniMax built it on the same design as its H3 video model, with the audio removed and a new image decoder added. The core model has 33.1 billion parameters, the internal settings a model learns in training and a rough guide to its size.

MiniMax's own figures say an image takes under 10 seconds, up to five times faster than mainstream rivals. Its model card calls it the best open model at following instructions. The card names no rival and shows no test scores, and no independent results had appeared by November 11.

Early interest was strong but shallow. One daily digest of trending models on November 11 listed it among five trending image, video and audio models, with 564 likes. It had been downloaded only 667 times.

Size helps explain the gap. The full model and the text-reading component it needs total about 135 gigabytes. One analysis found that a single top-end 80-gigabyte data center chip cannot hold both at full quality. Community-made compressed versions are much smaller.

The license is the business model

The license, dated November 10, does more than set a revenue limit. Commercial products must show the name MiniMax-H3-Image on screen. The model and its images may not be used to improve any other AI model, which blocks rivals from using its output as training material. Disputes go to Hong Kong courts.

The excluded regions are among the world's biggest markets for paid AI. Companies there can still apply to MiniMax for a license. MiniMax applied the same exclusions to its video model in August. A staff member then blamed legal and regulatory uncertainty, in a post on Hugging Face. In the US case, the staffer pointed to pending copyright lawsuits over AI video brought by Hollywood studios.

MiniMax also sells access through its own platform. List prices work out to about 2 cents for a standard square image and about 7 cents at full 16-megapixel size. A 50% launch discount on the largest images runs until November 17.

Running the model on your own hardware pays off only at large volumes. One analysis put the break-even point in the hundreds of thousands of images. Below that, the free download saves little.

The approach fits MiniMax's finances. The company listed in Hong Kong in January 2026. It reported 2025 revenue of about $79 million against a net loss of roughly $1.87 billion. Licensing fees and paid access are one way to narrow that gap.

Why a lower limit matters

The video model shows what MiniMax is protecting. H3 has passed 24 million downloads and more than 300 spin-off models, according to one industry report. Builders included an Nvidia and MIT research team, the video start-up Pika and the hosting firm fal. Nvidia's work reportedly runs under a separate licensing deal.

Comfy Org, which makes popular software for running image and video models, says it is the only official reseller of MiniMax commercial licenses. Its commercial package for running MiniMax models on a customer's own hardware starts at $5,000 a month. Its license page covers the H3 video model and audio models, not yet the image model.

Outside developers had also moved fast. A week before MiniMax's release, the AI-efficiency start-up Pruna launched its own image model built from the H3 video model. MiniMax's official version now competes with that and with established image tools from other firms.

The lower revenue limit widens the group of paying customers. Under the video license, a firm needed permission only once products built on the model earned $20 million. Under the image license, nearly any established agency, retailer or software company outside the excluded regions must talk to MiniMax first.

For executives and investors, the lesson goes beyond one company. Free-to-download no longer means free to use. Before building on any open model, check three things in the license. Where may it be used? What is the revenue cap? What are the limits on its output?