What you get

Meta released Llama 3 on April 18 under a document called the Meta Llama 3 Community License Agreement. The grant is generous on its face. You receive a non-exclusive, worldwide, royalty-free license to use, reproduce, distribute, and modify the Llama materials and to make derivative works. You own the derivative works you create. Nobody is charging you anything and there is no field-of-use limit beyond the acceptable use policy that travels with the agreement.

That is why people call it open. The word is doing a lot of work, though, because the same document contains four conditions that no license approved by the Open Source Initiative would carry. Two of them concern what you must say, one concerns how big you are allowed to be, and one concerns what you may do with the model's outputs.

The naming rules

If you distribute the model or a derivative, or a product or service that uses it, you must prominently display the phrase Built with Meta Llama 3 on a related website, user interface, blog post, about page, or product documentation. That is a trademark-style attribution embedded in a copyright license, and it applies to the product, not just to the weights.

The second naming rule is stricter. If you use the Llama materials to create, train, fine tune, or otherwise improve an AI model that you then distribute, you must include Llama 3 at the beginning of that model's name. A derivative model cannot be called by a fresh name with a footnote. Its name has to start with Meta's brand. Alongside that, any distribution must retain a notice file with the sentence stating that Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc. All Rights Reserved, and must ship the full agreement itself.

The output clause

The clause that matters most for research reads, in the license's words, that you will not use the Llama materials or any output or results of the Llama materials to improve any other large language model, excluding Meta Llama 3 or derivative works thereof. In plain terms, you may distil Llama 3 into a Llama 3 derivative, and you may not distil it into a model that is not one. Generating a synthetic instruction dataset with Llama 3 and training a Mistral or a from-scratch model on it is prohibited by the text.

Whether that clause is enforceable is a separate question and we are not going to pretend to answer it. The materials are weights, and the outputs are text the model produced for you. What we can say is that the restriction runs against the way the open model community actually works, where synthetic data from the best available model is the default way to bootstrap the next one. The license draws a line through that practice and puts Meta's own derivatives on the permitted side of it.

The 700 million user clause

The remaining condition is the one that had already appeared in the Llama 2 license. If the products or services of the licensee, or its affiliates, had more than 700 million monthly active users in the calendar month before Llama 3's release date, the grant does not apply and you must request a separate license from Meta, which Meta may grant or not at its discretion. The number is chosen so that a handful of very large companies are excluded and everyone else is not.

There are also the standard mechanics. The license is governed by California law. It terminates automatically if you breach it, at which point you must delete the materials. And if you start litigation against Meta alleging that the Llama materials or their outputs infringe your intellectual property, the license terminates as well. That last provision is common in patent-defensive open licenses. Here it also covers copyright claims about outputs.

What this means for reuse

For the foundation's own work, the practical reading is this. A fine-tuned Llama 3 released in the open must carry the Llama 3 prefix and the built-with credit, and that is easy to comply with. A synthetic dataset generated with Llama 3 is fine to release, but the license says it may only be used to train Llama 3 derivatives, which means the dataset's own license needs to say so or downstream users will violate a term they never saw. That second point is the one we expect to be widely missed, because a dataset on a hub page does not look like it carries Meta's terms.

We have not tried to catalogue which downstream projects comply with the naming rules and which ignore them, and we would not trust a quick survey to get that right. What we would want someone to do is track a sample of Llama 3 derivatives over the next year and record whether the prefix and the credit survive each hop of fine tuning and re-release. Our guess is that the credit erodes with distance from Meta, and a guess is all it is until someone counts.

Sources

  1. Meta, Meta Llama 3 Community License Agreement