Meta releases Llama 2 weights for commercial use, except to companies over 700 million monthly users at release

2023-07-18Industry

Meta, "Meta and Microsoft Introduce the Next Generation of Llama," ai.meta.com (July 18, 2023); Llama 2 Community License Agreement, Llama 2 Version Release Date July 18, 2023; Hugo Touvron et al., "Llama 2: Open Foundation and Fine-Tuned Chat Models," arXiv:2307.09288 (July 18, 2023)

Meta released Llama 2, a family of pretrained and fine-tuned large language models "ranging in scale from 7 billion to 70 billion parameters," with the weights and starting code downloadable and, in Meta's words, "free for research and commercial use." Meta published a research paper, a Responsible Use Guide and an Acceptable Use Policy, and named Microsoft its "preferred partner," with the models in the Azure catalog and also on Amazon Web Services and Hugging Face. The terms are set by the Llama 2 Community License Agreement, dated July 18, 2023. Section 1(a) grants a "non-exclusive, worldwide, non-transferable and royalty-free limited license" under Meta's "intellectual property or other rights" embodied in the Llama Materials to use, reproduce, distribute, copy, create derivative works of and modify them; the license does not mention patents by name. Section 1(b)(v) forbids using the Llama Materials or their outputs "to improve any other large language model" other than Llama 2 and its derivatives. Section 2 requires any licensee whose products had, on the Llama 2 version release date, "greater than 700 million monthly active users" in the preceding calendar month to request a separate license, granted in Meta's discretion. Section 5(b) leaves ownership of derivative works with the licensee, and Section 5(c) terminates the license on the date the licensee files any claim that the Llama Materials or their outputs infringe its intellectual property.

What it changesCommercially usable open weights from a frontier laboratory move the prior art baseline for every language-model application filed from today: a fine-tuning recipe, a prompting scheme or a retrieval pipeline that works on Llama 2 is now something a skilled artisan has the materials to try, and the examiner will treat the model as available. The patent program should respond in two ways. First, file on the fine-tuning methods, the data pipelines and the serving machinery the company builds on top of the base model, not on the base model's capabilities, because those are Meta's and are now public. Second, treat the license as a layer the in-house program tracks alongside patents: it grants no express patent license, it ends the moment the company sues Meta or anyone else claiming the model or its outputs infringe the company's intellectual property, and the "improve any other large language model" clause is a use restriction that a distillation or synthetic-data project will trip unless counsel is consulted first.

More on Trade secrets and open release · Prior art, public disclosure and obviousness

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.