Open source AI has a published definition: the Open Source Initiative's Open Source AI Definition 1.0, released on October 28, 2024, requires that users be free to use, study, modify, and share a system, and that its code, parameters, and detailed information about its training data be available under approved terms. Many models described as open release only their weights under custom licenses, which gives some of the benefits but not all of them. The advantages and uses of open source AI therefore depend on which kind of release a given model is.
Key takeaways
- Under the Open Source AI Definition, a system must provide code, parameters, and sufficiently detailed information about training data, with four freedoms: use, study, modify, and share.
- The Open Source Initiative has said Meta's Llama 3.x licenses are not open source, so "open" in a product name is not evidence of the definition being met.
- Ai2's OLMo 2 32B, released March 13, 2025, published its data, code, weights, and checkpoints, which lets outside researchers inspect and reproduce training steps.
- On one public leaderboard, the gap between the best closed-weight and best open-weight model fell from 8.04 percent in January 2024 to 1.70 percent in February 2025.
- The EU AI Act exempts some open-source AI from some duties, but not prohibited practices, high-risk systems, or models with systemic risk.
What the definition requires
The Open Source AI Definition adapts the idea behind open source software to systems that learn from data. It lists four freedoms: to use the system for any purpose without asking permission, to study how it works and inspect its components, to modify it for any purpose including changing its outputs, and to share it, modified or not. To make those freedoms real, the preferred form for making changes must be public. That means the complete code used to train and run the system, the model parameters such as weights, and enough information about the training data that a skilled person could build a substantially equivalent system. The data itself does not have to be handed over, but the description of it has to be detailed.
The standard matters because AI differs from ordinary software. A model's behavior depends on data that may be copyrighted, personal, or too large to publish, so releasing code alone says little about what the system does or why.
Three kinds of release compared
Public releases fall along a range. The table summarizes three documented examples and what each makes available; it describes the release, not the quality of the model.
| Release type | Example | What is public | Documented limit |
|---|---|---|---|
| Open weights under a custom license | Meta's Llama 3.x | Downloadable weights | The Open Source Initiative says the licenses fail its Open Source Definition on discrimination and field-of-use grounds (February 18, 2025) |
| Fully open model | Ai2 OLMo 2 32B | Pretraining and post-training data, training code, weights, and intermediate checkpoints | Benchmark claims come from the developer; the post says the model beat two commercial models on a set of academic tests |
| Open source library | scikit-learn | Source code for classical machine learning methods under a BSD license | A toolkit for building models, not a trained model |
What openness makes possible
Inspection is the clearest gain. Because Ai2 released OLMo 2's training data, code, and checkpoints, other researchers can study how a particular data choice changed behavior, instead of guessing from outside. That is the kind of work closed systems do not allow. The scikit-learn library shows an older version of the same pattern: it began in 2007 as a Google Summer of Code project, had its first public release on February 1, 2010, and is run by a community under a documented governance structure.
Cost and control are the other commonly cited benefits. Because the first freedom is the freedom to run a system for any purpose, an organization can run an open system on its own hardware, adapt it to its own documents, and avoid a dependency on one vendor's pricing and terms. These gains are conditional: running a large model takes hardware and staff, and the sources reviewed here do not publish cost comparisons that would settle which route is cheaper.
Performance has also moved. Stanford's 2025 AI Index reports that on the Chatbot Arena Leaderboard the leading closed-weight model beat the top open-weight model by 8.04 percent in early January 2024, and that by February 2025 the gap had narrowed to 1.70 percent. The figure compares open-weight models, which are a broader group than open source AI, and it covers one human-preference ranking.
Where the label misleads
The Open Source Initiative has used the term open washing for products that borrow the open source name without the freedoms. In a post dated February 18, 2025, it cited the Free Software Foundation's review of the Llama 3.1 community license, which found failures on the freedom to use the software for any purpose and on two points of the Open Source Definition. A model can be free to download and still be restricted in who may use it and for what.
A second gap is data. Weights alone do not reveal what a model was trained on, so bias, copyright, and privacy questions stay unanswerable. A third is risk: openness lets anyone inspect a system and also lets anyone modify it, so the balance of benefit and harm depends on the model and its use.
How the EU AI Act treats open-source AI
The EU's Regulation 2024/1689 contains two relevant provisions. Article 2(12) says the regulation does not apply to AI systems released under free and open-source licenses, except systems that are high-risk, fall under the prohibited practices in Article 5, or fall under the transparency rules in Article 50. For general-purpose AI models, Article 53(2) exempts providers of models under a free and open-source license from certain documentation duties if the weights, architecture information, and usage information are public, but not if the model carries systemic risk. The European Commission states that obligations for general-purpose AI model providers began applying on August 2, 2025, and that its enforcement powers started on August 2, 2026. This is general information, not legal advice, and a provider's actual duties depend on its facts.
Checking a claim of open source AI
- Is the license one the Open Source Initiative has approved, and does it limit who can use the model or for what?
- Are training code and parameters public, or only a downloadable file?
- Does the developer describe the training data in enough detail for someone to rebuild a similar dataset?
- Who maintains the project, and how are changes decided?
- Do performance claims come from independent evaluations or from the developer's own post?
- Does a law such as the EU AI Act still apply despite the open license?
FAQ
Is an open-weight model the same as open source AI?
No. Open weights means the parameters can be downloaded. Open source AI under the Open Source Initiative's definition also needs code, detailed data information, and licenses that grant the four freedoms.
Does open source AI mean free to use for any business?
Only when the license says so. Some widely downloaded models carry community licenses with limits on users or fields of use, and those limits are the reason the Open Source Initiative does not count them.
Is open source AI safer or riskier than closed AI?
The sources reviewed do not support a general answer. Openness allows outside inspection and also allows modification, so the risk depends on the model, the use, and who has access.
Are there official statistics on open source AI use at work?
None turned up in the sources reviewed for this article. Adoption figures for open versus closed models come from developers and research groups, so they should be read as such.
Related reading
- How organizations apply AI in practice
- What deep learning models do
- Computer vision as an AI application
- OpenAI's structure and products
More in AI and Society.
Last verified: October 2026
Sources
- The Open Source AI Definition 1.0, Open Source Initiative, October 2024
- Meta's Llama license is still not Open Source, Open Source Initiative, February 18, 2025
- OLMo 2 32B: First fully open model to outperform GPT 3.5 and GPT 4o mini, Allen Institute for AI, March 13, 2025
- About scikit-learn, scikit-learn project, 2026
- Technical Performance, AI Index Report 2025, Stanford Institute for Human-Centered AI, 2025
- Regulation (EU) 2024/1689, Articles 2 and 53, European Commission AI Act Service Desk, 2026
- Guidelines for providers of general-purpose AI models, European Commission, 2026