Meta's Llama licence added a naming rule between versions

A free licence with a ceiling

Meta distributes its Llama models under a "community licence" rather than a licence approved by the Open Source Initiative. The Llama 2 Community License Agreement, published when Llama 2 was released in July 2023, grants broad rights to use, reproduce and create derivative works, but with a ceiling: "If, on the Llama 2 version release date, the monthly active users...is greater than 700 million...you must request a license from Meta". Below that threshold, a company can build on the model without paying a fee; above it, the free grant no longer applies and a separate commercial agreement is required.

What changed for Llama 3

The Meta Llama 3 Community License Agreement, dated to the model's release on 18 April 2024, keeps the same 700-million-user threshold but adds a requirement the Llama 2 text did not contain: a distributed AI model built from the Llama Materials must carry "Llama 3" "at the beginning of any such AI model name", and redistributors must display "Built with Meta Llama 3". The Llama 2 licence instead asked only for a copyright notice in a separate "Notice" text file; it did not require the word "Llama" to appear in a derivative model's own name. Both versions prohibit using outputs to train a competing large language model other than Llama itself.

A licence, not a definition of open source

These are contractual terms set by one company, not a neutral standard: the threshold, the naming rule and the training restriction can all change again in a future version, and Meta - not a court or standards body - decides what counts as a qualifying derivative. Under the Open Source Initiative's 2024 Open Source AI Definition, a user-count ceiling and a use restriction on training competing models are each, on their own, enough to place a licence outside that definition, regardless of whether the model's weights are freely downloadable.

• Does the deploying organisation's monthly active user count, or its parent company's, approach the 700 million threshold in either version?
• If shipping a fine-tuned or distilled derivative, does its name and any required attribution match the version-specific naming clause?
• Has the specific licence text been re-read for the version actually being deployed, rather than assumed unchanged from an earlier Llama release?

A licence that reads as permissive at a glance can still carry a scale trigger and a naming obligation that only a full read of the current version's text will reveal. Terms changed once already between Llama 2 and Llama 3, and nothing in either document promises they will not change again.