Glossary

Open-source AI model

Written with AI from the sources below and checked against our facts sheet. Last checked 7 Nov 2026. How this site is made · Spot an error?

Part of: EU AI Act

An open-source AI model is released under terms that let anyone use, study, modify and share it. Under the Open Source Initiative's definition, that means publishing its weights, code and training-data details. Many "open" models publish only weights under restrictive licenses. The EU AI Act gives free and open-source models partial exemptions, never for high-risk systems or systemic-risk models.

In plain English

Think of a cake. Open weights is getting the finished cake: you can serve it, slice it and add icing. Open source is also getting the recipe, the ingredient list and the baker's notes, with permission to bake and sell your own version.

The Open Source AI Definition 1.0 from the Open Source Initiative sets four freedoms: use, study, modify and share, for any purpose. It also requires the "preferred form" for changing the model: data information, code and parameters. Many models marketed as open publish only the weights, under licenses that restrict use.

Example

Harlow Health, a fictional hospital network, needs a clinical-notes summarizer that runs on its own servers. It shortlists three options.

Model	What is released	Problem
A	Proprietary API only	Patient data would leave Harlow's servers
B	Weights under a community license	Bans medical use and needs a separate license above a user threshold
C	Weights, code and data documentation under Apache 2.0	None found in review

Harlow picks Model C. Legal confirms the license permits commercial fine-tuning. Harlow then runs its own red teaming, writes its own model card and sets up patching and monitoring. Open weights moved that work from a vendor to Harlow.

Why it matters

EU AI Act. Two exemptions apply, both partial.

AI systems (Art. 2(12)). The Act does not apply to AI systems under free and open-source licenses. The exceptions are high-risk systems and those under Art. 5 or Art. 50.
General-purpose AI models (Art. 53(2)). The license must allow access, use, modification and distribution, and the weights, architecture and usage information must be public. Such models skip the two documentation duties in Art. 53(1)(a) and (b). The copyright policy and training-content summary still apply. Models with systemic risk get no exemption.

Product liability. The new EU Product Liability Directive excludes free and open-source software supplied outside a commercial activity.

Licenses. Model licenses fall into four families: permissive (MIT, Apache 2.0), copyleft (GPL, AGPL), responsible-AI (RAIL and similar) and non-commercial. Responsible-AI licenses usually require you to pass their use restrictions on to your own customers.

How the AIGP exam tests it

Model types are covered in lesson IV.A.2, licenses in lesson IV.B.4 and the GPAI exemption in lesson II.C.8.

Common traps:

"Free to download, free to use." Many licenses restrict commercial use or set conditions.
Open weights equals open source. Not under the OSI definition.
Exemption overreach. No exemption for systemic-risk models, high-risk systems, Art. 5 or Art. 50.
Fine-tuning washes off restrictions. Responsible-AI licenses typically carry their terms into derivatives.
Closed means secure. An absolute; security depends on operation, not secrecy.

The tell: the governance question is never "is it open?" It is "what does this license let us do, and what does it make us do?"

Frequently asked questions

What is the difference between open source and open weights?

Open weights means the trained parameters are published, so anyone can download and run the model. Open source, under the Open Source Initiative's definition, goes further. It requires the weights, the full training and inference code, and enough information about the training data to rebuild a substantially equivalent system. It must also come with freedom to use, study, modify and share.

Are open-source AI models exempt from the EU AI Act?

Partly. AI systems under free and open-source licenses are outside the Act (Art. 2(12)). The exceptions are high-risk systems, Art. 5 practices and Art. 50 transparency cases. Open-source general-purpose AI models with public weights skip two documentation duties (Art. 53(2)). They still need a copyright policy and a training-content summary, and systemic-risk models get no exemption.

Can I use an open-source AI model commercially?

It depends on the license. Permissive licenses such as Apache 2.0 allow commercial use if you keep the notices. Many model licenses add use restrictions, attribution rules or user-count thresholds, and some forbid commercial use entirely. Downloadable weights never guarantee commercial rights, so read the license before building on the model.

Are open-source AI models less safe than closed models?

Not inherently. Security depends on how a model is built, tested, hosted and operated, not on whether its weights are secret. Open weights allow outside inspection and self-hosting, but the deployer takes on security, patching and safety testing that a vendor would otherwise handle. Either choice can be safe or unsafe.

Sources

Spot an error? Tell us