Open Weight Is Not Open Source

Update, 23 October 2026. We re-read Meta’s Llama licences and use policies and OSI’s definition, approved-licence list and statements on Llama, and reworked the page around the question readers ask: is Llama open source? One correction: we said none of the Llama restrictions reach someone running a model at home. The acceptable use policy does, because it binds anyone who uses the model. We also wrote as if every Llama licence requires “Built with Llama”; the Llama 2 licence does not. The 700 million user clause, the attribution line in 3.3 and the Open Source Definition’s criteria are unchanged from our September read, and Meta’s licence pages have moved to dev.meta.ai. The table by generation, OSI’s verdicts, the EU limit and the questions at the end are new.

No: Meta’s Llama models are not open source by the Open Source Initiative’s definition, and no Llama licence is on OSI’s list of approved licences. OSI has said so twice, about Llama 2 in July 2023 and Llama 3.x in February 2025: the licences fail its criteria 5 and 6, which bar discriminating against people or groups and restricting fields of endeavour.

Llama is open weight. Meta publishes the weights and lets you use, modify, redistribute and build commercial products on them, on its own terms. Above 700 million monthly active users, those terms need a separate licence from Meta, which Meta may grant “in its sole discretion”.

Llama’s licence, generation by generation

	Llama 2	Llama 3.3	Llama 4
Licence date	18 July 2023	6 December 2024	5 April 2025
Licence from Meta needed above 700 million monthly active users	yes	yes	yes
Acceptable use policy binds every use	yes	yes	yes
Multimodal models withheld from people and companies based in the EU	no	yes	yes
“Built with Llama” shown on a product or site	no	yes	yes
“Llama” at the start of a model trained from it	no	yes	yes
Bars using it to improve other large language models	yes	no	no
On OSI’s approved list	no	no	no
OSI’s own verdict	fails points 5 and 6 (2023)	fails points 5 and 6 (2025, Llama 3.x)	no OSI statement found

From Meta’s licence and use-policy pages and OSI’s list, read 23 October 2026. The 700 million figure is counted in the calendar month before each version’s release date, across the products of the licensee and its affiliates. The EU limit is in the use policies for Llama 3.2, 3.3 and 4; it does not apply to end users of a product built on those models.

Why OSI says Llama is not open source

The Open Source Definition is ten criteria, and two of them decide this. Criterion 5: “The license must not discriminate against any person or group of persons.” Criterion 6: “The license must not restrict anyone from making use of the program in a specific field of endeavor.”

OSI’s 2023 statement applies both to Llama 2. The licence “puts restrictions on commercial use for some users (paragraph 2)”, which is the 700 million clause, “and also restricts the use of the model and software for certain purposes (the Acceptable Use Policy).” Its 2025 statement says Llama 3.x is “still not Open Source by any stretch of the imagination”, fails the same two points, and adds a restriction “excluding any persons in the European Union”.

The two clauses fail for different reasons. The user ceiling decides who gets the licence by the size of their business, not by what they do with the model. The use policy decides what the model may be used for: Meta’s list includes military applications, and OSI points to two other listed areas, regulated substances and critical infrastructure, as uses that “might be highly beneficial to society”.

None of that makes Meta’s terms unreasonable. It makes “open source” the wrong label, and the label is what people rely on when they choose what to build on.

Does OSI’s AI definition change the answer?

No. The Open Source AI Definition 1.0 applies the same test to models. Its first freedom is to “Use the system for any purpose and without having to ask for permission”, and it requires the weights to be available under “OSI-approved terms”. It also asks for training code and enough information about the training data to build a substantially equivalent system, which the software definition does not. OSI’s 2025 statement says there is “no need” for the newer definition to judge Llama: the licence fails the older one first.

What it means if you run Llama at home

The use policy still applies. Meta’s use policy binds anyone who runs the model, at home included. Meta’s policy for Llama 3.3 opens: “If you access or use Llama 3.3, you agree to this Acceptable Use Policy”, and the licence makes following it a condition of use. Its prohibited list is about illegal activity, malware, weapons, fraud and similar harms rather than ordinary chat, writing or coding help.

The rest attaches to what you share. Passing on the licence, the “Built with Llama” line and the naming rule apply when you “distribute or make available” the model, or a product or another AI model containing it. Llama 2 instead bars using it or its outputs to improve another large language model.

If you want a model whose licence is on OSI’s list, the Qwen3-32B model card states Apache 2.0, and Qwen 3.6’s two open models carry the same licence. Which of them runs well is a question of your graphics card, and setting one up locally takes three steps.

Before you build on an open-weight model

• Read the licence file that came with the weights, not an article about the family. Llama 2 and Llama 3.3 differ on attribution, model naming and training other models.
• If you are in the EU and want a multimodal Llama, read that version’s use policy first.
• Choose the licence before the benchmark. Swapping models late because of a clause costs more than picking one whose licence fits from the start.
• Treat “open source” in a model announcement as a claim to check. OSI’s list is public, and checking a licence against it takes a minute.

Questions people ask

Is Meta Llama open source?

Not by the Open Source Initiative’s definition. Llama is open weight: Meta publishes the weights under its own Llama Community License, which OSI says fails its definition because it restricts who may use the model and for what. No Llama licence is on OSI’s approved list.

Is the Llama license OSI approved?

No. OSI’s list of approved licences, read 23 October 2026, has no Llama or Meta licence on it. A licence gets there only through OSI’s review process, and OSI has said publicly that the Llama 2 and Llama 3.x licences do not meet its Open Source Definition.

Is the Meta Llama 2 license open source or not?

Not open source, according to OSI’s statement of 20 July 2023. It said the licence restricts commercial use for some users, through the 700 million user clause, and restricts some purposes through the acceptable use policy, failing points 5 and 6 of the Open Source Definition.

Is Llama 4 open source?

Its licence is not on OSI’s approved list, and it keeps the two clauses OSI objected to in Llama 2 and 3.x: a licence from Meta above 700 million monthly active users, and an acceptable use policy. Its use policy also withholds multimodal models from people and companies based in the EU.

What is the difference between open weight and open source?

Open weight means the trained weights are published for download, on whatever terms the maker sets. Open source means the licence meets OSI’s definition, including use by anyone for any purpose. OSI’s AI definition also asks for training code and data information.

Can I use Llama commercially?

Yes, within Meta’s terms. The licence allows commercial use, but you must follow the acceptable use policy, pass on a copy of the licence when you distribute it, and for Llama 3.3 and 4 display “Built with Llama”. Companies whose products had over 700 million monthly active users need Meta’s permission.

Does the Llama licence apply if I only run it at home?

Partly. The acceptable use policy applies to anyone who uses the model, at home included. The “Built with Llama” line and the duty to pass on the licence apply only when you distribute the model or make a product containing it available.