Meta gated LLaMA behind a non-commercial research licence

A paper and a gated release, three days apart

Meta's announcement, dated 24 February 2023, introduces LLaMA as a family of foundation models in '7B, 13B, 33B, and 65B parameter' sizes, released under 'a noncommercial license focused on research use cases'. Access was granted case by case, to academic researchers and affiliated government, civil-society and industry labs that applied through a form linked from the paper. The accompanying paper, submitted three days earlier on 27 February 2023, states the goal directly: training smaller models on more tokens than is compute-optimal by the Chinchilla rule, so that the resulting models are cheaper to run at inference even if more expensive to train.

What the paper claims, and under what training scale

The abstract states that 'LLaMA-13B outperforms GPT-3 (175B) on most benchmarks', and that the 65B version 'is competitive with the best models, Chinchilla-70B and PaLM-540B'. These are the paper's own comparisons, on the benchmarks and prompting settings it selected, trained on 'trillions of tokens' using only what the authors describe as publicly available datasets. Meta's blog repeats the parameter-count line-up without repeating the comparison figures themselves, and instead emphasises access and responsible use. A benchmark table under one paper's chosen conditions is not the same claim as a general ranking, and the report does not claim its own smaller models are more capable in any absolute sense, only more efficient per parameter served.

A licence, not a technical access control

The gated process Meta describes controls who receives an approved copy of the weights on request; it does not by itself control what happens to a copy once distributed, because a software licence is a contract enforceable after the fact rather than a lock on the file. Meta's post acknowledges upfront that LLaMA 'shares the challenges' of prior models, including 'bias, toxicity, and hallucinations', and argues that controlled research access would let the field study those problems collectively, which presumes the recipient list stays the effective boundary on who can run the model. The current repository, as retrieved on 16 September 2026, notes that the original release now lives in a separate branch and describes a different, later access process for subsequent Llama versions; it should be read as a living document about today's repository, not a record of the 2023 terms.

Questions to carry into your own evaluation

• Does a 'smaller model matches a larger one' claim specify the benchmarks, prompting format, and whether the comparison model was itself instruction-tuned?
• Is the licence for the weights you are using the one the paper describes, or a later replacement with different terms?
• Does controlled access to weights meaningfully limit downstream use once a copy has left the approved distribution channel?

LLaMA's research release is best read as two separate claims: a training-efficiency argument made under stated benchmark conditions, and a licensing choice that assumed access could be controlled. The paper's comparisons hold only within the terms it sets; the licence held only as long as the distribution channel did.