Llama Community License

What the license grants

The Meta Llama 3 Community License is the agreement Meta attaches to the model files it publishes for Llama 3. Under it, a licensee receives a non-exclusive, worldwide, non-transferable and royalty-free limited license to use, reproduce, distribute, copy, create derivative works of and modify what the document calls the Llama Materials^[1]. This version of the agreement carries the release date 18 April 2024^[1].

Each of those adjectives does work. Non-exclusive means Meta keeps the right to grant the same permissions to everyone else, so a licensee gets no exclusivity over the model. Royalty-free means no fee is owed for the use itself, however much of it there is. Non-transferable means the permission belongs to the licensee and cannot simply be handed on to a third party like an asset. And "limited" is the important one: every permission in the document is conditional, and those conditions are what make this license unusual.

The conditions

The best-known condition is a size threshold. A licensee with more than 700 million monthly active users in the preceding calendar month must ask Meta for a license, and Meta may grant it at its sole discretion; until it does, that licensee is not authorized to exercise the rights in the agreement^[1]. The practical effect is that one document gives different rights to different users: for a small developer the license is a standing permission, for a very large platform it is a request that can be refused.

Two attribution duties follow. A distributor must prominently display "Built with Meta Llama 3" on a related website, user interface, blog post, about page or product documentation. And if the materials are used to train or improve another AI model that is then distributed, that model's name must begin with "Llama 3"^[1]. The second rule ties a derived model's public identity to the original — a naming requirement rather than a fee.

The license also restricts what the materials may be used for. It prohibits using the Llama Materials or their outputs to improve any other large language model, except Meta Llama 3 or its derivatives^[1]. A permissive open source license would allow exactly that, so this clause is the clearest point of departure. Finally, the Acceptable Use Policy is incorporated by reference, so the effective terms are two documents rather than one, and the license terminates automatically if a licensee institutes litigation against Meta alleging that the materials infringe the licensee's rights^[1].

Where the disagreement comes from

Whether this counts as open source is contested. TechCrunch reports that OSI's Stefano Maffulli has been openly critical of Meta calling its models open source, and that after discussions with OSI, Google and Microsoft agreed to drop the term for models that are not fully open, while Meta has not^[2]. Meta took part in the OSAID drafting process but disagreed with the result: a spokesperson said there is no single open source AI definition, and that Llama's license and acceptable use policy help keep people safe by having some restrictions in place^[2]. TechCrunch also notes that Meta is one of the companies funding the OSI's work^[2].

What it means in practice

For most users the license behaves like a permissive one: download the weights, run them, fine-tune them, ship a product. The terms start to matter at the edges — a platform large enough to cross the user threshold, a team that wants to use Llama outputs to train a competing model, or a company that wants to publish a derived model under a name of its own choosing. In those cases the answer comes from the license text rather than from the general expectation that open weights can be used for anything.

The document is best read as an intermediate category: more permissive than a proprietary API, more conditional than a standard open source license, with the conditions written so that they travel with the model and stay visible to end users through the attribution requirements.