Llama 2 shipped open weights under a custom licence

Open weights, released with a named partner

Meta's announcement, dated 18 July 2023, introduces Llama 2 as weights and code 'free for research and commercial use', released in partnership with Microsoft, which the post names as 'the preferred partner for Llama 2' with availability through Azure's model catalog alongside AWS and Hugging Face. The accompanying paper, submitted the same day, reports pretrained and fine-tuned models from 7 billion to 70 billion parameters, with the fine-tuned Llama 2-Chat variants said to 'outperform open-source chat models on most benchmarks' and to compare favourably with some closed models in human evaluations the authors ran themselves.

What 'free for commercial use' actually permits

The stated freedom is bounded by the Llama 2 Community License, which is not the Apache or MIT licence the 'open-source' framing might suggest. Two clauses carry the licence's substantive restrictions: a company with more than '700 million monthly active users' at the product's release date must separately 'request a license from Meta, which Meta may grant... in its sole discretion', and no licensee may 'use the Llama Materials or any output or results of the Llama Materials to improve any other large language model' other than Llama 2 itself or its derivatives. The licence also terminates automatically against anyone who brings a patent claim against Meta over the materials. None of these are conditions attached to a conventional open-source licence, which by definition does not discriminate against fields of use or against named categories of user.

Why the qualifier matters more than the weight release

Meta's paper documents its own safety process, including 'internal and external red-teaming' and a published responsible-use guide, and frames open access as enabling wider scrutiny of the model. That argument depends on 'open' meaning inspectable and runnable, which the weight release does deliver; it does not depend on the licence being free of use restrictions, and the Community License carries several. A reader who treats 'open' and 'open source' as interchangeable will miss both the training-data restriction, aimed squarely at competitors training on Llama 2's outputs, and the scale threshold that exempts everyone except a small number of the largest technology companies.

Questions to carry into your own evaluation

• Does 'open' in a given announcement mean the weights are downloadable, the licence imposes no field-of-use restriction, or both?
• Would the deployment you are planning fall under the 700-million-user carve-out, or under the restriction on using outputs to train a competing model?
• Do the paper's human-evaluation comparisons to closed models specify who ran the evaluation and on what prompt set?

Llama 2's release is accurately described as open weights under a permissive-looking but custom licence, not as open source in the sense the term carries in software more generally. The distinction is not pedantic: it determines who is legally free to build on the model, and for what.