AI model licences: MIT, Apache 2.0, Llama, Gemma… what each one allows

Being able to download an AI model doesn’t mean you can do anything you like with it. Every model comes with a licence. Some, like MIT or Apache 2.0, have almost no conditions; others ban commercial use or cap the number of users. Here are the most common licences and the licence each well-known model uses, checked against the official texts.

Checked on 16/10/2026 against the official text of each licence and each model’s official card. Licences change: reread the model card before you use a model.

Educational content, not legal advice. For a specific commercial use, read the full licence text and, if needed, ask a lawyer.

Table 1: what each licence allows

“With limits” means commercial use is allowed except in the cases in the restrictions column. “OSI-approved” was checked against the official SPDX list and opensource.org.

Licence families in a nutshell

• Permissive (MIT, Apache 2.0, BSD). You can use, modify and sell, as long as you keep the copyright notice. Apache 2.0 also grants you a licence to the author’s patents and asks you to state what you changed.
• Copyleft (GPL, AGPL). If you distribute a modified version, you must release it under the same licence, with its source. The AGPL extends this to programs you offer over the internet. They’re rare for model weights and more common in the software that runs them.
• Creative Commons (CC BY, CC BY-NC, CC BY-SA). Designed for content and data, not software. BY requires attribution, NC bans commercial use and SA requires derivatives to be shared under the same licence. They’re common for datasets.
• Responsible AI licences (OpenRAIL, OpenRAIL-M). They allow use and sale but include a list of prohibited uses that you must pass on to anyone who receives the model. The “M” means the licence applies to the model.
• Each company’s own licence (Llama, Gemma, Qwen, Mistral, FLUX…). Each sets its own rules: user caps, acceptable use policies or a ban on commercial use. None is OSI-approved.

Table 2: which licence each model uses

One family can have different licences depending on the version or size. That’s why Qwen, DeepSeek, Mistral, Falcon, Gemma and FLUX appear on several rows.

Closed models: GPT, Claude and Gemini

These models can’t be downloaded: you use them through an app or an API, and what governs their use is each company’s service terms, not a weights licence.

The key clauses, word for word

Verbatim quotes from the official English texts, which are the binding ones.

• Llama 4 Community License, §2: “If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta”
• Llama 3.1 / 4, §1.b.i: “prominently display “Built with Llama” on a related website, user interface, blogpost, about page, or product documentation”
• Llama 4 Acceptable Use Policy (Llama 3.2: same clause): “With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union. This restriction does not apply to end users of a product or service that incorporates any such multimodal models.”
• Gemma Terms of Use, §3.2: “Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement.”
• Gemma Terms of Use, §3.3: “Google claims no rights in Outputs you generate using Gemma.”
• Qwen License Agreement, §4: “If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, you shall request a license from us.”
• Qwen Research License, §1.i: ““Non-Commercial” shall mean for research or evaluation purposes only.”
• DeepSeek Model License, §5: “You cannot use the Model and the Derivatives of the Model for the specified restricted uses.”
• MRL 0.1, §3.2: “You shall only use the Mistral Models, Derivatives (whether or not created by Mistral AI) and Outputs for Research Purposes.”
• TII Falcon License 1.0: “use-based restrictions incorporating the Acceptable Use Policy in the manner set out in Section 5 of this license”
• Stability AI Community License: “If at any time You or Your Affiliate(s), either individually or in aggregate, generate more than USD $1,000,000 in annual revenue”
• FLUX.1 [dev] License, b): “You may only access, use, Distribute, or create Derivatives of the FLUX.1 [dev] Model or Derivatives for Non-Commercial Purposes.”
• FLUX.1 [dev] License, d): “You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein.”
• Modified MIT (Kimi K2): “if the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display “Kimi K2” on the user interface”
• CreativeML OpenRAIL-M: “the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license”
• Apache 2.0, §3: “each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license”
• OSI, Open Source AI Definition 1.0: “Data Information: Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.”

Open weights is not the same as open source

A model with open weights only means you can download its trained parameters. It doesn’t mean you know what data it was trained on or that you get the training code.

The Open Source Initiative (OSI), which decides what counts as open source, published the Open Source AI Definition (OSAID 1.0). It requires the freedoms to use, study, modify and share, plus three pieces: the parameters, the complete code and enough information about the training data to rebuild an equivalent system.

So a model whose weights are under Apache 2.0 may not meet the OSAID if it doesn’t publish information about its data. And licences with user caps or prohibited uses, such as Llama’s, are not on the OSI-approved list, even when presented as “open”.

What to check before using a model in your business

1. The exact version. Look up the licence of that specific model and size on its official card; it can change within a family.
2. Whether commercial use is allowed. Look for words like “non-commercial”, “research” or “non-production”.
3. User or revenue caps. Llama (700 million monthly users), Qwen (100 million), Stability (USD 1 million annual revenue), Kimi K2.
4. Where you are. Llama’s multimodal models (Llama 3.2 Vision and Llama 4, which Meta describes as natively multimodal) exclude individuals domiciled in, and companies with their principal place of business in, the EU, though not their end users.
5. The acceptable use policy. Many licences incorporate one and can update it.
6. What you must display or include. A copy of the licence, notices, “Built with Llama” or the model name.
7. Outputs. Gemma and FLUX.1 [dev] expressly claim no rights over what you generate, but FLUX.1 [dev] bans using it to train a competing model.
8. Data and privacy. The licence doesn’t cover the GDPR or the EU AI Act: those are separate.

Quick glossary

• Permissive licence: Allows use, modification and sale with few conditions, usually keeping the authorship notice.
• Copyleft: Requires modified versions you distribute to carry the same licence.
• Attribution: Crediting the original author and keeping their notices.
• Share-alike: What you build from the work must carry the same licence.
• Weights: The model’s learned parameters: the file you download.
• Derivative work: A model modified or fine-tuned from the original.
• Acceptable use policy: A list of prohibited uses that the licence incorporates.
• OSI: Open Source Initiative, the organisation that approves open source licences.
• SPDX: A standard list of licence identifiers, such as “Apache-2.0”.

More terms in the glossary: open weights, open source, fine-tuning.

What we couldn’t verify

Falcon 180B doesn’t declare its licence in its Hugging Face card metadata, so we left it out. For closed models we only link the terms: we don’t summarise them because they change often.