Using open AI models in your organisation: what the licences allow
Most open AI models are released under Apache 2.0 or MIT and may be used, modified and redistributed in your organisation without conditions. Some vendors use their own licences: Meta, for example, requires a name notice, prohibits certain uses and excludes companies based in the EU for fourth-generation multimodal models. What counts is always the licence file of the specific model version, because licences change within a model family.
What licences do open AI models come with?
Open models make their weights available for download. The licence sets the conditions under which you may use them. Broadly, there are three groups (as of October 2026):
| Licence | Examples | Commercial and public-sector use | Conditions |
|---|---|---|---|
| Apache 2.0 | open models from OpenAI, Alibaba, most from Mistral AI, the new generation from Google | permitted | include the licence text and notices when redistributing, mark changes |
| MIT | models from various vendors, also tools such as Ollama | permitted | include the licence text when redistributing |
| Vendor’s own licence | Meta, older models from Google, some models from Mistral AI | mostly permitted, with conditions | name notice, usage prohibitions, thresholds by size or revenue, in some cases exclusion of regions |
Apache 2.0 and MIT are licences approved by the Open Source Initiative. They permit use, modification and redistribution for any purpose and exclude warranty and liability. Apache 2.0 additionally includes a patent licence from the contributors.
What do the vendors’ own licences say?
Meta releases its models under its own community licence. For the fourth generation (effective 5 April 2025), the following applies, among other things:
- Name notice: Anyone who redistributes the models or a product using them must display “Built with” and the model name prominently. Anyone who uses the models or their answers to train another model and redistributes it must begin that model’s name with the name of the model family.
- Large providers: From 700 million monthly active people, a separate licence from Meta is required.
- Prohibited uses: The acceptable use policy rules out, among other things, operating critical infrastructure and the unauthorised practice of healthcare professions.
- EU exclusion for multimodal models: For multimodal models (text and image) of this generation, individuals domiciled in the EU and companies with their principal place of business in the EU are not granted the usage rights. End customers of a product that contains such a model are exempt. A hospital or public authority in the EU that runs the model itself, however, is not an end customer in this sense.
- Law: The law of California applies, and the courts there have jurisdiction.
Google had its own terms of use with a list of prohibited uses for its open models up to the third generation. There, Google reserved the right to restrict use “remotely or otherwise” if it believed there had been a violation. The new generation from April 2026, by contrast, is under Apache 2.0.
Mistral AI releases most of its open models under Apache 2.0. According to the company, some are under a modified MIT licence: companies with more than 20 million US dollars in monthly revenue then need a commercial licence or use the model via the Mistral AI platform. Other models are only released for research or non-production purposes.
What should hospitals and public authorities watch out for?
- Version, not family: The licence can change from one generation to the next, as with Google. Check the licence file of the specific version you use and record it with the date.
- Usage prohibitions: Lists of prohibited uses are broadly worded, for example medical activity without a licence to practise or critical infrastructure. Whether your use falls under them should be clarified before rollout, especially in clinical settings and critical infrastructure (KRITIS).
- Regions: An exclusion for companies based in the EU also affects municipalities and hospitals that run the model themselves.
- Foreign law: Vendors’ own licences often refer to US law and US courts. That does not always fit the requirements of public contracting authorities.
- Tools are not models: Software such as Ollama is under MIT; the models it loads each have their own licences. The tool’s licence says nothing about the model’s.
- Redistribution: If you pass a model on to third parties, for example to other organisations in a group or as a service provider, the redistribution obligations apply: licence text, notice file and, with Meta, the name notice.
Is an open model open source within the meaning of the AI Act?
The AI Act relieves AI systems and models under a “free and open-source licence” in some places (Art. 2(12), Art. 53(2)). According to recitals 102 to 104, this requires that the licence freely permits use, modification and redistribution and that the model is not provided against payment. Whether licences with usage prohibitions or thresholds for large companies meet this is disputed. For organisations that only use a model, this changes little: their obligations as deployers depend on the use, not on the model’s licence. In its 2024 definition, the Open Source Initiative additionally requires information on training data and code.
How does this work with basebox?
basebox selects the AI model and runs it, in our cloud or on your server. A model register documents the developer, version and licence of each model. If your organisation has requirements on the origin of a model, we implement them. Your own models with an OpenAI-compatible interface can be connected; you then check their licence yourself. More on proof under Security.
Not legal advice: This guide reflects the state of our research on 11 October 2026. Licences change with every model version. What counts is the licence text of the version you use. If in doubt, have it reviewed by a lawyer.
Frequently asked questions
Is an open model the same as open source?
Not necessarily. Open usually only means that the weights are available for download. Under the Open Source Initiative's definition, information on training data and code is also required; licences with usage prohibitions or thresholds for large companies do not meet it.
Do the model's answers belong to the vendor?
Not under the licences mentioned here. Google, for example, states explicitly that it claims no rights in generated content. Whether the answers themselves are protected by copyright is a separate question; it is governed by copyright law, not by the model licence.
Do we need a licence if we only use the model internally?
Yes, the licence also applies to internal use. Apache 2.0 and MIT set no conditions for it. Vendors' own licences can also apply internally, for example through lists of prohibited uses or the exclusion of certain countries.
Sources
- Apache License 2.0 ↗
- MIT License (Open Source Initiative) ↗
- Meta: community licence for the fourth model generation (effective 5 April 2025) ↗
- Meta: acceptable use policy for the fourth model generation ↗
- Google: terms of use for older open models (as of 1 April 2026) ↗
- Google: new generation of open models under Apache 2.0 (April 2026) ↗
- Mistral AI: under which licence are the open models available? (12 August 2026) ↗
- Mistral AI: licences of the model weights (documentation) ↗
- OpenAI: open models under Apache 2.0 (August 2025) ↗
- Alibaba: open models under Apache 2.0 ↗
- Open Source Initiative: Open Source AI Definition 1.0 ↗
- Regulation (EU) 2024/1689 (AI Act), Art. 2(12), Art. 53(2), recitals 102 to 104 (EUR-Lex) ↗
- Ollama: licence (MIT, GitHub) ↗
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 11 October 2026.