Does the EU AI Act exempt open-source AI models? Not the way most people assume. The exemptions in Article 2(12) and Article 53(2) of Regulation (EU) 2024/1689 are narrow, they attach to the open-source system or the model rather than to the company that runs it, and the duties that matter most stay in place.
Stand: 2026-09-29
The line comes up in almost every first conversation: "We run local open-source models, so the AI Act does not concern us." It sounds reasonable, because the Act does name open source in several places and does promise relief. The relief just lands somewhere other than where most people expect. Treating it as a blanket exemption buys a false sense of safety that the text does not support.
The mistake that gets expensive
The AI Act draws a consistent line between the provider, who develops and supplies a model or system, and the deployer, who puts it into use under their own authority. The open-source exemptions are written for the provider. A company that runs Llama, Qwen or Mistral on its own hardware is, in almost every case, a deployer, not a provider. For that role the Act sets out no general open-source carve-out.
That is the real trap. The model's licence says nothing about which duties your business carries. Whether a model is open, partly open or proprietary decides its maker's obligations. Your own obligations turn on what you use the model for and which risk tier that use falls into. Keep those two questions apart, or you will draw a conclusion about your own position from the freedom of the model, and the text does not license that leap.
What Article 2(12) actually exempts
Article 2(12) takes AI systems released under a free and open-source licence out of the scope of the Regulation. The clause that follows makes the exemption narrow. It does not apply where the system is placed on the market or put into service as a high-risk system under Article 6 and Annex III, nor to the prohibited practices under Article 5, nor where the system is subject to the transparency obligations of Article 50.
That leaves little behind. What survives those three exceptions to the exemption are AI systems of minimal risk, a spam filter or a spell checker, say. Such systems carry no real obligations under the Act to begin with, open source or not. Lawyers therefore call this exemption largely declaratory: it states expressly for open source what already holds for minimal-risk systems in general, but it lifts no concrete duty that would otherwise have applied. Read Article 2(12) as a free pass for a CV-screening tool or a credit-scoring system and you have skipped the decisive clause, because those are precisely the high-risk uses it excludes.
What Article 53(2) spares the model provider, and what it does not
The second exemption, and the one that matters more in practice, concerns general-purpose models, the large language models themselves. Under Article 53(1) a provider of a general-purpose model owes four things: technical documentation of the model under point (a), information for downstream providers under point (b), a policy to comply with EU copyright law under point (c), and a sufficiently detailed summary of the training content under point (d).
Article 53(2) relieves a provider who releases the model under a free and open-source licence of the first two, the technical documentation and the downstream information. The other two remain. The copyright policy under point (c) and the public training-content summary under point (d) apply to an open model as well. This is the part most often overlooked: an open model must still set out what it was trained on and how it respects copyright.
The exemption also carries conditions. The licence has to allow access, use, modification and distribution, and the parameters, including the weights, the architecture and information on model use, have to be publicly available. If the model is offered for a fee, the exemption falls away. A model with weights behind a paywall is not open source for the purposes of this provision, whatever the word "open" in its licence suggests. And the most important limit sits at the end: for models with systemic risk under Article 55, in practice the frontier models trained above 10^25 FLOPs of compute, the exemption does not apply. They carry the full obligations irrespective of licence. A related point affects providers established outside the EU: under Article 54(6) they need not appoint an authorised representative in the Union for an open-source model without systemic risk.
The point almost everyone misses
Everything so far concerns the provider of the model. For an operating business the decisive question is a different one: what applies to me when I deploy an open model? The answer is not in the exemption provisions. It sits with the deployer duties, and those know no open-source carve-out.
Since 2 February 2025, Article 4 has required every deployer to take measures to support the development of AI literacy among the people working with the system. Since 2 August 2026, the transparency duty under Article 50 has applied: a chatbot must disclose that it is a machine, and AI-generated content must be marked as such. Deploy a high-risk system and the Article 26 duties join from 2 December 2027, a date set by Regulation (EU) 2026/1744 of July 2026. None of these duties asks which licence the model carries. Where your own role tips from deployer to provider, for instance through a substantial modification of the model, we set out in deployer or provider.
What this means for a business running a local open model
In practice it comes down to a short check that pays off before you lean on the licence. Classify the use, not the model, first. A local model for internal document search is minimal risk; the same model used to pre-screen job applications is high-risk under Annex III, and the open licence changes nothing there. Then settle transparency: if the system talks to customers or produces published content, Article 50 applies. Third, see to the AI literacy of the people involved, the most easily overlooked and most easily met duty of the set. Finally, keep data protection and the AI Act apart: the GDPR applies in parallel and knows no open-source question at all.
The advantage of an open model lies elsewhere, and it is real. Holding the weights yourself and running the model on your own hardware gives you full control over the data, frees you from dependence on any provider contract, and lets you evidence the model's provenance. That is a concrete contribution to data sovereignty and makes every later audit easier. It is an advantage in meeting the duties, though, not a release from them. EU funding instruments back the actual project, not the licence of the model.
Open source is an advantage, not a free pass
The short version for the board: the EU AI Act rewards open source, but it does not exempt you. Article 2(12) and Article 53(2) ease the model maker's position on tightly drawn points, copyright and training data stay mandatory even there, and models with systemic risk are excluded entirely. For your business as a deployer, what counts is not the licence but the purpose of use. Draw that line cleanly and you spare yourself the unpleasant surprise at the first audit.
Which risk tier your specific use falls into, and which duties follow from it, is best settled against the actual project. That is exactly what we do in a pilot project before a model goes into regular operation. This text sets out the legal position based on our reading and does not replace individual legal advice.
Frequently asked questions
Does the EU AI Act exempt open-source AI models from all obligations?
No. Article 2(12) exempts open-source AI systems only while they are not prohibited, not high-risk and not caught by the transparency rules. Article 53(2) relieves the provider of an open-source general-purpose model of two documentation duties, and nothing more.
If I only use an open-source model, do I still have obligations?
Yes. The exemptions apply to the provider of the model, not to you as the deployer. Anyone who puts a model into operational use carries the deployer duties under Article 4 and Article 50 regardless of the licence, plus Article 26 for a high-risk system.
What stays mandatory for an open-source model provider despite the exemption?
Even under a free licence, Article 53(1)(c), the copyright policy, and Article 53(1)(d), the public summary of training content, still apply. For a model with systemic risk under Article 55, the exemption does not apply at all.
Since when have these rules applied?
The obligations for general-purpose models have applied since 2 August 2025, enforceable through the Commission's powers since 2 August 2026. The Article 50 transparency duty applies since 2 August 2026, and the high-risk deployer duties from 2 December 2027.