Open-weight model licenses for game developers, compared

Every license claim below was checked against the model owner’s own license page or license file on the date at the top of this article. Open-weight licenses get rewritten without much warning, sometimes attached to a new model in the same family with no announcement beyond a changed word in a HuggingFace repo. Treat the checked date as the edge of what this page can promise you, and read the actual license for the exact checkpoint you plan to ship before you rely on it. Reading tool terms before you ship covers the reading skills this article assumes.

Is “open weights” the same thing as “open source”?

No, and the difference is the entire reason this article exists. Open source has a real definition: no restriction on who can use the software, what for, or how much money they make doing it. Every family below fails that test somewhere. Llama bars you above a user count. Qwen does the same at a different threshold. Gemma’s older versions carry a Prohibited Use Policy that follows the model downstream even after you redistribute it. “Open weights” means you can download the numbers and run them yourself. It says nothing about whether you are allowed to ship what you build with them. Read the license, not the word “open” on the landing page.

Can you ship a commercial game on Llama?

Yes, under the Llama 4 Community License Agreement, and Meta’s terms have kept the same shape since Llama 2. Below 700 million monthly active users on the products you or your affiliates ship, you owe nothing beyond compliance. Above that line, you need a separate agreement with Meta before you can use the model at all, which will not touch an indie team but matters if a publisher or platform holder is involved. Two conditions attach regardless of size: you must display “Built with Llama” if you distribute the model or a product built on it, and if you fine-tune or train a model that you then distribute, its name has to start with “Llama.” An Acceptable Use Policy also applies. One more thing worth knowing: Llama 4’s multimodal (vision) weights cannot be used by or distributed to individuals or companies domiciled in the EU; the text-only path is unaffected.

Can you ship a commercial game on Mistral?

For the current generation, yes, cleanly. Mistral Large 3, Mistral Small 4, and the Ministral 3 family are Apache 2.0, no threshold, no attribution requirement, no acceptable-use flow-down. The catch is that “Mistral” is not one license. Codestral sits under a commercial-only Premier license rather than being freely redistributable, and older or specialty checkpoints, including Mistral Large 2 and Pixtral Large, remain under the Mistral Research License, which is non-commercial. If you picked a Mistral model more than a year ago, its license may not match what a fresh download of the same family name gets you today.

Can you ship a commercial game on Qwen?

Yes, and for the current generation, size does not change the answer. Qwen3-generation checkpoints are Apache 2.0 across the range, including the large flagship mixture-of-experts models, with no threshold and no attribution requirement. The catch is generational rather than size-based: some older Qwen2.5-generation checkpoints, including Qwen2.5-72B, still carry the separate Qwen License Agreement, which permits commercial use for free below 100 million monthly active users and requires requesting a license from Alibaba Cloud above that line. If a project or tutorial points at a Qwen2.5 model rather than Qwen3, check that specific repo’s LICENSE file instead of assuming the current generation’s Apache default carries backward.

Can you ship a commercial game on Gemma?

It depends which Gemma. Gemma 4, released in April 2026, is Apache 2.0: no Prohibited Use Policy, no notice requirement, no restriction beyond the standard license. Gemma 3 and the specialty variants built on it (PaliGemma, ShieldGemma, EmbeddingGemma) still run on the older custom Gemma Terms of Use, which permits commercial use but attaches a Prohibited Use Policy that you must pass along to anyone you redistribute the model to, plus a notice requirement and a right for Google to restrict access if it believes you are violating the terms. If you already have a Gemma pipeline built before April 2026, it is worth checking which generation it is actually running.

Can you ship a commercial game on Stable Diffusion?

Under the Stability AI Community License, yes, for free, as long as your organization’s annual revenue from any source stays under one million US dollars. Cross that line and you need a paid Enterprise license regardless of whether the revenue came from the game itself. The Community License also requires you to display “Powered by Stability AI” if you ship a product built on it, and to carry a specific attribution notice if you redistribute the weights or a derivative. Older checkpoints, SD 1.5 and SDXL, run on the earlier CreativeML Open RAIL-M license instead, which has no revenue threshold but bakes a list of prohibited uses directly into the license text rather than a separate policy document. Stable Diffusion 3.5 is the flagship checkpoint under the Community License as of this check; confirm you are reading the license attached to the specific weights you download, since Stability has changed license terms on a new release before without changing the family name.

Can you ship a commercial game on FLUX?

Only on specific variants, and the family name tells you nothing on its own. FLUX.1 [schnell] and the newer FLUX.2 [klein] 4B are both Apache 2.0, fully commercial, no threshold. Every [dev] checkpoint, at both FLUX.1 and FLUX.2, ships under a non-commercial license instead: free for testing and research, not for a production pipeline, with the added restriction that you cannot use its output to train a competing model. The larger FLUX.2 [klein] 9B also landed non-commercial despite early announcements suggesting otherwise, and [pro], [flex], and [max] are proprietary, API-only products with their own commercial terms. If a tutorial tells you to grab “the FLUX model” without naming the bracket, assume it means dev and check before you build a pipeline around it.

This table is a starting point, not a substitute for the license text. Licenses in this space have moved at least once a year across every family listed here.

Going deeper

Two habits carry past this table. First, a fine-tune inherits its base model’s conditions; fine-tuning Llama 4 on your own dataset does not detach the result from the 700M MAU cap or the naming requirement, and fine-tuning a FLUX [dev] checkpoint does not make the result commercial. Second, watch for license chains: a game-dev tool that embeds one of these models internally is bound by that model’s license, and its own terms to you sit on top of that, not instead of it. If a tool’s terms are silent on which base model it runs, that silence is itself worth asking about before you build a project around the tool.

If a specific checkpoint’s terms still read ambiguous, ask about it in the AI Game Dev Org Discord before you commit a project to it, not after.

Sources

• Llama 4 Community License Agreement
• Mistral AI models overview and licenses
• Qwen License Agreement (legacy 72B-class checkpoints)
• Gemma Terms of Use
• Stability AI Community License
• Black Forest Labs open-weights licensing