Reading the Llama 4 Community License line by line

What shipped

Meta released Llama 4 on April 5 with two downloadable models and one announced. Scout has 17 billion active parameters across 16 experts, about 109 billion total, and a claimed 10 million token context. Maverick has the same 17 billion active parameters across 128 experts, about 400 billion total, and a 1 million token context. Behemoth, at 288 billion active parameters and roughly 2 trillion total, was still training at launch. Both released models take text and images as input, output text, cover 12 languages, and have an August 2024 knowledge cutoff.

The weights come with the Llama 4 Community License Agreement and an Acceptable Use Policy. Most people click through both. We read the license on the Scout model card so that we could say precisely what it allows, and this post walks the clauses that decide whether you can use it and what it costs you in obligations if you do.

The 700 million user clause

The clause everyone has heard of reads, in the Scout license text, that if on the Llama 4 version release date the monthly active users of the products or services made available by or for the licensee are greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta. Meta may grant it or not at its discretion. In practice this excludes a handful of companies by name without naming them, and the date anchor matters. The test is taken on the release date, so a company that crosses 700 million users next year is not caught retroactively.

For a research foundation or a startup the clause is irrelevant in effect. The reason it matters to the rest of us is definitional. The Open Source Initiative objected to this clause in Llama 2 back in 2023, on the grounds that discriminating against a class of users violates the Open Source Definition, and nothing about it has changed across three major versions. Any license that says some users need permission is a source-available license with a commercial carve-out. It can be a generous one, and this one is, and it is still not open source by the definition the term has had for over two decades.

Branding and naming obligations

Two clauses impose obligations that scale with success rather than with size. First, if you distribute the Llama materials or a product that contains them, you must prominently display Built with Llama on a related website, user interface, blog post, about page, or product documentation. Second, if you use the Llama materials or their outputs to create, train, fine-tune, or otherwise improve an AI model that you distribute, you must include Llama at the beginning of that model's name.

The naming rule is the more interesting of the two. It applies to outputs as well as weights, which means a model trained on synthetic data generated by Llama 4 is, on the license's terms, a Llama-named derivative. That is a broad reach, and it is a reach the license claims through contract rather than copyright, since the copyright status of model outputs is unsettled. Whether the clause would hold up if tested is a question for lawyers. What it does today is make every downstream project either carry Meta's brand or take a legal risk, and most carry the brand.

You also have to retain a specific attribution notice in any distribution. The Scout card gives the wording, which reads that Llama 4 is licensed under the Llama 4 Community License, Copyright Meta Platforms, Inc., All Rights Reserved. Those last three words sit oddly in something described as open, and they are accurate about the legal posture.

Acceptable use and the EU

The license requires compliance with applicable laws including trade compliance laws, and adherence to the Acceptable Use Policy. The policy is a separate document that Meta can update, and the license incorporates it by reference. Prior versions restricted uses touching controlled substances and critical infrastructure among others, and the OSI cited field-of-endeavour restrictions of this kind as a second reason the Llama licenses fail the definition, since open source licenses are not permitted to restrict what you use the software for.

The Wikipedia summary of the license history also records that later versions restricted use by any individual in the European Union, and the Free Software Foundation, in January 2025, classified Llama as nonfree, citing both the popular-application restriction and the enforcement of trade regulations outside the user's jurisdiction. We have not seen Meta state a rationale for the EU carve-out on the Scout card itself, so we will only note that the restriction exists and that it is a further discrimination against a class of users.

What you can actually do

Set the objections aside for a moment and read the license as a user. Below 700 million users, you can download the weights, run them commercially, fine-tune them, distribute the fine-tunes, and build products on them, provided you brand, name, attribute, and stay inside the use policy. That is more than most commercial model providers allow and it is why so much of the open-weight ecosystem for the last two years has been Llama-shaped. The license is a good deal. It is just not the deal the word open describes.

One line from a researcher quoted in the Wikipedia article puts the other half of the problem well. There is no source to be seen, and the training data is entirely undocumented. Even with a permissive license, weights without data and code are a binary. You can run it and modify it, and you cannot rebuild it or audit what went into it. For our purposes at the foundation that is the bigger gap. A license we could live with would not fix a release we cannot reproduce.

What we would like to see next is simple. Either Meta adopts a license that passes the definition, and then the argument is over, or the field settles on a term for what this actually is. Open weights is honest. Open source is not, and each release under the current terms makes the word mean a little less.