"Open source" has a settled meaning in software. For AI models the term is used loosely, and the loose use causes real surprises. Many releases that people call open source publish only the model's weights, which are the trained parameters. Those releases can still come with restrictions on who may use them and how. This guide separates the terms and shows what to look for.
Three things people mean
- Open weights. The trained parameters are available to download and run. That says nothing on its own about training data, training code or the terms of use.
- Source available / community licence. The weights, and sometimes more, are available under a custom licence that grants some freedoms and withholds others.
- Open source AI, as defined by the OSI. The Open Source Initiative published the Open Source AI Definition 1.0, which sets requirements for a system to qualify.
What the Open Source AI Definition requires
The definition grants four freedoms: to use the system for any purpose without permission, to study how it works and inspect its components, to modify it for any purpose including changing its output, and to share it with or without modifications.[1]
To make those freedoms practical, it asks for three kinds of material:[1]
- Data information: "Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system." This covers a description of the data, its provenance, scope and characteristics, how it was obtained and selected, labelling procedures and processing methods.
- Code: "The complete source code used to train and run the system," including data processing and filtering code, training code with its arguments and settings, validation and testing code, supporting libraries, and inference code.
- Parameters: "The model parameters, such as weights or other configuration settings," made available under OSI-approved terms, including checkpoints from intermediate training stages and the final optimizer state.
Note what this means for a release that includes only weights. The weights are one of three components, and the parameters must also be offered under OSI-approved terms.
A worked example: the Llama 4 licence
Meta publishes Llama models under a custom licence. The Llama 4 Community License Agreement, effective 5 April 2025, includes several conditions worth reading closely.[2]
| Clause | What it says (in summary) |
|---|---|
| 700 million monthly active users | If, on the Llama 4 version release date, monthly active users are greater than 700 million, the licensee "must request a license from Meta" (the licence defines whose users count, so read the full clause) |
| "Built with Llama" notice | The licence requires providing a copy of the agreement and prominently displaying "Built with Llama" (see the text for when this applies) |
| Model naming | For AI models built with Llama materials, the licence requires "Llama" at the beginning of the model name (see the text for exactly which models this covers) |
| Acceptable Use Policy | Use must follow applicable laws and the Acceptable Use Policy, which the licence incorporates by reference |
Those are Meta's words as published for Llama 4. Other Llama versions and other model families have different licences, so read the one that ships with the model you intend to use.
Our reading
The first freedom in the OSI definition is the freedom to use the system "for any purpose without permission." A licence that requires permission above a user threshold, or that requires a specific notice and name, imposes conditions that the definition does not describe. We would not call such a release open source AI under the definition. That is our interpretation of the two documents, and the Open Source Initiative or a lawyer may reason differently, so read both texts yourself.
None of this makes a community licence a bad choice. For most teams, the conditions above may never bind them. The point is to know them before you build.
A short process before you adopt a model
- Read the licence text itself, not a summary or a model card blurb.
- List what your product needs: commercial use, redistribution, fine-tuning, use of outputs, expected user scale.
- Check each need against the licence, and flag clauses tied to scale, naming or field of use.
- Record the licence version you relied on, because terms change between model releases.
- Ask counsel about anything ambiguous. This article is not legal advice.
If you are also deciding how a model connects to your systems, see what Model Context Protocol is, and for how to judge a vendor's claims about a model, how to read an AI benchmark claim.