What is LLaMA 2 AI?

On 18 July 2023, Meta released Llama 2 alongside a technical paper whose title deserves a slow read: "Llama 2: Open Foundation and Fine-Tuned Chat Models." Open. That same day, anyone who wanted to download the model's weights had to accept a licence containing two clauses that the canonical definition of "open source" does not allow.

This is neither a scandal nor a trick: it is the industry norm. And understanding it is worth considerably more than memorising the model's parameter count. If you are building a product on an "open" model, or you want to know what your company actually chose the day it chose one, what you may and may not do is written in a contract of roughly six thousand words — not in the announcement headline.

What Llama 2 is, using the published figures

The paper's abstract puts it plainly: Llama 2 is "a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters." The family shipped in three sizes — 7B, 13B and 70B — and in two flavours: the base models and Llama 2-Chat, tuned for dialogue.

In that same abstract Meta states that its models "outperform open-source chat models on most benchmarks we tested" and that, based on its human evaluations for helpfulness and safety, they "may be a suitable substitute for closed-source models."

Look closely at the exact shape of that sentence, because it is the first usable thing in this piece: the party doing the measuring is the party doing the selling. "Most benchmarks we tested" and "our human evaluations" are honest formulations precisely because they hedge: a vendor who writes "we tested" is admitting, without saying so, that it picked the tests. That does not invalidate the result. It places it. And the gap between reading "beats the open models" and reading "beats the open models on the tests I chose" is the whole distance between believing and understanding.

"Open weights" is not "open source"

What Meta released are the weights: the millions of numbers that come out of training and that constitute the working model. That is a great deal, and it is real. Anyone can download them, run them on their own hardware, inspect them, fine-tune them for their own case and — this is the genuinely unusual part — use them commercially at no cost. Against a model reachable only through a paid API, the difference is enormous.

But "open source" is not an intensity adjective, like "fast" or "powerful," that each party stretches as far as suits them. It is a term with a written definition, maintained by the Open Source Initiative since 1998, and the interesting part is that you can check it in five minutes.

The two clauses, in their own words

The Llama 2 licence, published in full in Meta's official repository, contains a section headed "Additional Commercial Terms." It reads, verbatim:

"If, on the Llama 2 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee's affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion."

And a few paragraphs earlier, among the use restrictions:

"You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof)."

Now the two conditions from the Open Source Definition to weigh them against. The fifth: "The license must not discriminate against any person or group of persons." The sixth: "The license must not restrict anyone from making use of the program in a specific field of endeavor."

The 700-million-user threshold carves out a group — a very small one, and not by accident: it is the size of Meta's direct competitors — that needs individual, discretionary permission. The ban on improving other language models restricts one specific field of endeavour. Both are perfectly legitimate as business decisions, and both are incompatible with that definition.

What the announcement summary never tells you

Reading the whole licence, rather than its summary, pays. There are at least three things inside that almost nobody mentions and that change real decisions.

The first corrects a widespread misreading: the 700-million threshold is not inherited. Section 1.b.ii states that "if you receive Llama Materials, or any derivative works thereof, from a Licensee as part of an integrated end user product, then Section 2 of this Agreement will not apply to you." In other words: the clause that frightens half the industry reaches only those who download and distribute, not those who consume the model inside somebody else's product.

The second is an active, quiet obligation. Anyone redistributing the materials must hand a copy of the agreement to the other party and must retain, in a text file named Notice, this exact attribution: "Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved." The licence grants no trademark rights either: using the name is permitted only to the extent "reasonable and customary" in describing the material.

The third matters most to anyone shipping to production. The warranty disclaimer covers not just the model but what the model produces: the materials "and any output and results therefrom" are provided "AS IS," without warranties of any kind, and "you are solely responsible for determining the appropriateness of using or redistributing" them and assume the risks. Translated: if the model generates something that lands you in trouble, the trouble is yours. This is effectively identical across almost every model, open and closed, and it is worth having read once.

The capability: three questions for any licence that calls itself open

This is what stays with you once Llama 2 is history. Faced with any model advertised as open, open the licence file — it is always there, and almost nobody opens it — and look for three things:

One: may I use the model's output to train or improve another model? This is the clause that decides whether the model is a tool or a dead end. If the answer is no, that model cannot be the starting point for anything of your own that competes with it.

Two: is there a threshold that changes the regime? Users, revenue, sector, country. A threshold means the licence you read is not the licence that will apply to you if you succeed. Better to know that before succeeding than after.

Three: who decides the borderline cases, and with how much latitude? The phrase to hunt for is "sole discretion." Where it appears, the other party is under no obligation to give you reasons or to treat you the same as anyone else.

Those three questions will get you through any model licence in the time it takes to drink a coffee. And they are needed, because the pattern has not stopped repeating: nearly every vendor that publishes weights writes its own bespoke licence, with its own threshold and its own carve-outs, and nearly all of them use the word "open" in the announcement.

Where to go next, with no middlemen

The material is complete and free. The Llama 2 technical paper includes the description of the fine-tuning and the safety work. The full licence is six thousand words, reads in twenty minutes, and holds everything summarised here. The acceptable use policy is incorporated into the licence by reference — meaning it binds you just the same, even though it lives on another page — and it is a good example of something always worth checking. The Open Source Definition is ten points and fits on one screen.

The conclusion is not that Llama 2 is worse than advertised. It is that a model with published weights and a restrictive licence remains one of the most useful things anyone has handed the industry, and that calling it by its proper name — open weights, bespoke licence — takes nothing away from it. The capability you take from this is reading the licence before the headline, and knowing which three questions to ask it.