Llama 3’s License: What the Terms Actually Permit

Commercial use of Llama 3 is permitted, and that is the shortest true answer. The longer one is that permission comes with attribution obligations, a naming rule, a notice file and one threshold clause that applies to almost nobody but is the reason the licence is not open source.

It is not an open source licence

The Meta Llama 3 Community License Agreement was published with the Llama 3 release on 18 April 2024. It is a custom licence written by Meta, not Apache 2.0, not MIT, and not approved by the Open Source Initiative. It fails the Open Source Definition on the same point every time: it discriminates between users, both through the user-count threshold below and through the use restrictions in the acceptable use policy it incorporates by reference.

“Open weights” is the accurate term and it is a different claim from “open source”. The weights are downloadable and usable commercially, which is the part most people need. The licence is nevertheless a contract with conditions, and it can be breached. The full text is at Meta’s Llama 3 licence page, with the accompanying acceptable use policy published alongside it.

For contrast in the same territory, Mistral releases several of its models under Apache 2.0 with no equivalent obligations — what Apache 2.0 means for a weights release is the page for that comparison, and the difference in obligations is the substance of it.

The 700 million user threshold

This is the clause the licence is known for. Section 2, “Additional Commercial Terms”, provides that if, on the version release date, the products or services made available by or for the licensee or its affiliates had more than 700 million monthly active users in the preceding calendar month, the licensee must request a separate licence from Meta, which Meta may grant or withhold at its sole discretion.

Three details about how that is drafted matter more than the number itself:

• It is measured on the version release date, not continuously. Crossing 700 million users after the release does not retroactively trigger it for that version — though the next version you adopt is measured on its own release date.
• It counts affiliates. A small subsidiary of a very large group is measured against the group.
• It does not forbid use; it requires asking. The obligation is to request a licence, and Meta’s discretion is explicit.

For virtually every reader this clause is inert. It is included here because it is the specific provision that makes the licence non-open-source, and because a legal review will ask about it whether or not it applies.

What you must do if you redistribute

Running the model inside your own product carries very light obligations. Distributing the model, or a derivative of it, carries real ones. The Llama 3 licence requires all of the following:

1. Pass the agreement on. Provide a copy of the community licence with any copy of the Llama Materials you distribute.
2. Display the attribution string. Llama 3 requires “Built with Meta Llama 3” to appear prominently on a related website, user interface, blog post, about page or product documentation.
3. Include the notice file. A text file distributed with the materials must contain: “Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.”
4. Name derivatives correctly. A model you distribute that is a derivative of Llama 3 must carry “Llama 3” at the beginning of its name. This is why fine-tunes on Hugging Face are named the way they are; it is a licence obligation, not a convention.
5. Comply with the acceptable use policy, which is incorporated by reference and which Meta may update.

The licence also disclaims warranties and limits liability in the usual terms, provides for termination on breach, and is governed by California law. It does not require you to publish your fine-tuning data, your prompts or your application code — there is no copyleft here.

The ordinary case: running it in a product

Most readers are not distributing weights. They are serving a model behind their own product and want to know what they owe. The list is much shorter, and the shape of it is worth stating explicitly because the redistribution obligations above are often applied to situations that do not trigger them.

Serving inference from a model you host, where users receive generated text and never the weights, is not distributing the Llama Materials. The clauses about the notice file and the derivative-naming rule attach to distribution. What still applies is the acceptable use policy, which governs use rather than distribution, and the attribution requirement, which Meta words in terms of a related website, user interface or documentation — so a line in your about page or docs saying the product is built with Llama is the low-cost way to be clearly inside it.

Three boundaries are worth being careful about, and all three are places where a product quietly becomes a distributor:

• Shipping the model to the client. An on-device model in a mobile app, or weights downloaded into a desktop application, is distribution. The 1B and 3B models exist for exactly this use case, so it is not a hypothetical.
• Publishing a fine-tune. Uploading adapters or merged weights anywhere public brings in the naming rule and the notice file, whatever the size of the change.
• Exposing the model as a model. Offering an inference API over Llama to third parties sits closer to the redistribution provisions than serving an application does, and the attribution and naming obligations are the ones to read carefully.

What changed after 3.0

The community licence is re-issued with each release and the text is not identical. Two changes are worth knowing because they reverse advice that was correct at the time it was written.

• Improving other models. The Llama 3 licence of April 2024 prohibited using the Llama Materials or their outputs to improve any other large language model, excluding Llama 3 and its derivatives. The Llama 3.1 licence of 23 July 2024 removed that prohibition and permits using outputs to train other models, subject to the naming rule below. Synthetic data generation and distillation moved from forbidden to permitted between two point releases three months apart.
• The naming and attribution strings. Llama 3.1 generalised the attribution to “Built with Llama” and requires that any AI model you distribute that was created, trained or improved using Llama Materials or their outputs carries “Llama” at the beginning of its name. Note the widened scope: it now reaches models trained on Llama outputs, not only derivatives of the weights.

A third change belongs to Llama 3.2 specifically: its licence withholds the grant for the multimodal models from individuals domiciled in, and companies with a principal place of business in, the European Union. That restriction and its scope are covered in the Llama 3.2 vision models page. The Llama 3.1 text of the licence is at Meta’s Llama 3.1 licence page and the 3.2 text at the Llama 3.2 licence page.

Reading the licence you are actually under

Because the terms move between versions, “the Llama licence” is not a thing you can be compliant with in general. Four practical consequences:

• Record which version you accepted. The gated download flow makes you accept a specific agreement for a specific release. Keep a copy with the checkpoint, not a link — the page can change.
• Re-read on every version bump. Upgrading from 3.1 to 3.3 is accepting a different contract. The synthetic-data change above is the proof that this is not a formality.
• A quantised or fine-tuned re-upload is still bound. A community GGUF of a Llama checkpoint is a derivative and carries the same obligations, whatever the uploader wrote in the repository description. Third-party licence metadata is not authoritative.
• The permission is a contract, not a property of the file. This is the point where licence terms and what deprecation means for open weights meet: nobody can stop the weights running, and that is a separate question from whether you are permitted to run them.