Meta Llama

Open weights, and a licence that is not open source

Llama models are downloadable. You can fetch the weights, run them on your own hardware, fine-tune them, and serve them without asking anyone.

That is not the same as open source, and the difference is where most of the surprises live. The terms are the Llama 4 Community License Agreement, effective 2025-04-05, and it is a custom licence written by Meta rather than Apache 2.0 or MIT. A team that reads "open weights" and assumes the permissions they know from other open-source projects will be wrong in four specific ways.

Four conditions to read before you build on it

Taken from the agreement, checked 2026-09-16.

Attribution in your product. You must display "Built with Llama" prominently on a related website, interface, or documentation, and reproduce the notice: "Llama 4 is licensed under the Llama 4 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved."

Naming. Any AI model you create using the materials must have a name beginning with "Llama". This is the condition that catches teams who planned to fine-tune and ship the result under their own product name. You can do the fine-tune; the model has to carry the prefix.

A user threshold. An organisation with more than 700 million monthly active users must request a separate commercial licence from Meta and may not use the materials without it. See the next section for why that clause exists.

Acceptable use. A separate Acceptable Use Policy is incorporated into the agreement by reference, so it binds you the same as the text you read.

None of these are onerous for most teams. All of them are invisible if you assume a standard licence and never open the file.

What the 700 million threshold is doing there

A user-count threshold in a software licence is unusual enough to be worth explaining, because understanding it tells you how to read the rest.

The clause is not aimed at you unless you are one of a handful of companies. It exists to keep the largest platforms from taking the weights and competing directly, while leaving everyone smaller free to build. Meta gets distribution and an ecosystem; the companies that could most cheaply turn the weights against it have to come and ask.

Read that way, the licence is a commercial instrument rather than a community one. That is not a criticism, and it is the right frame for the other clauses: attribution and naming exist so that Llama's presence in the ecosystem stays visible, which is the return Meta is collecting.

Version discipline matters here. The licence is versioned alongside the model generation, so "the Llama licence" is not one document. Check the agreement that ships with the weights you are downloading.

What you take on by running the weights yourself

The licence is the part specific to this family. Everything else is the ordinary cost of self-hosting, which self-hosted models covers in full: weights are the easy number, the KV cache grows with concurrency, and running out preempts requests that then get recomputed.

Two decisions are yours the moment you download rather than call an API.

Quantization. The published weights are one precision, and most people run something smaller. That trade is real and adjustable, and Ollama covers how packaging tools make the choice for you if you do not make it yourself.

A serving runtime. Weights are a file. Turning them into an endpoint that handles concurrent requests is a separate piece of software with its own behaviour under load.

What you get for that work is the reason to do it: no per-token bill, no rate limit, no third party receiving your prompts, and a model that does not change underneath a tested prompt. That last one is worth more than it looks, because a hosted model has a retirement date and this one does not.

Further reading

• Meta, Llama 4 Community License Agreement: the attribution, naming, threshold and acceptable-use terms quoted above.
• Self-hosted models: the hardware and cost side of running open weights.
• Ollama: the fastest way to run one of these locally.
• Qwen: the other open-weight family covered here.