When Open-Source Models Are Exempt From GPAI Obligations

The open-source carve-out in the AI Act is narrower than its reputation. It removes two of the four Article 53 obligations, it depends on conditions about the licence and the published artefacts that many “open” releases do not meet, and it switches off entirely for models with systemic risk.

The conditions in Article 53(2)

Article 53(2) of Regulation (EU) 2024/1689 disapplies the obligations in Article 53(1), points (a) and (b), for providers of AI models released under a free and open-source licence that allows for the access, usage, modification and distribution of the model, and whose parameters — including the weights, information on the model architecture and information on model usage — are made publicly available. The text is on EUR-Lex.

Unpacked, that is two cumulative conditions, and each has teeth:

• The licence must permit all four freedoms. Access, usage, modification and distribution. A licence that forbids commercial use, that caps use above a user threshold, or that restricts redistribution does not obviously satisfy this, whatever it is called in the announcement.
• Three artefacts must be publicly available. The parameters including the weights, information on the model architecture, and information on model usage. Publishing weights on a hub with no architecture description and no usage information is not the full condition. Note also what is not on the list: training data, training code and the training recipe are nowhere in Article 53(2).

What the exemption does not remove

This is the part most summaries get wrong. Article 53(2) names points (a) and (b) only:

• Point (a) — disapplied. Drawing up and keeping up to date technical documentation of the model, including its training and testing process, for the AI Office and national competent authorities on request (Annex XI).
• Point (b) — disapplied. Making information and documentation available to downstream providers who integrate the model into their AI systems (Annex XII).
• Point (c) — still applies. The copyright policy, including identifying and complying with Article 4(3) CDSM reservations. See the copyright policy page.
• Point (d) — still applies. The publicly available, sufficiently detailed training content summary, following the AI Office template. See the training content summary page.

So an open-weights release still owes a copyright policy and a published training content summary. That is a coherent design once you see the logic: (a) and (b) are about supplying information to authorities and to integrators, which openness substantially achieves by itself, whereas (c) and (d) protect third parties whose interests openness does nothing for.

Separately, Article 2(12) excludes AI systems released under free and open-source licences from the Regulation altogether — but with exceptions that swallow most of the interesting cases: it does not apply where the system is placed on the market or put into service as a high-risk AI system, or where it falls under Article 5 (prohibited practices) or Article 50 (transparency). Article 2(12) is about systems; Article 53(2) is about models. Conflating them is the most common error in this area.

Where it stops: systemic risk

The second sentence of Article 53(2) is unambiguous: the exception shall not apply to general-purpose AI models with systemic risk. An open-weights model above the Article 51(2) presumption threshold of 10²⁵ floating point operations of training compute, or designated by the Commission under Article 51(1)(b), owes the full Article 53(1) set and the Article 55 obligations described on the systemic-risk obligations page.

This is where the exemption stops being a plausible strategy for a frontier lab and remains a real one for everyone else. Article 55 includes duties — adversarial testing, incident tracking, cybersecurity protection of the model — that are awkward to discharge for a model whose weights you have deliberately handed to everybody. How a provider evidences model-weight security for a model published openly is not resolved by the text, and it is one of the more interesting unanswered questions in Chapter V.

The monetisation question

Recital 103 describes free and open-source AI components as those shared openly where users can freely access, use, modify and redistribute them, and it treats monetisation as taking a component outside that characterisation. Article 53(2) itself contains no monetisation condition — it speaks only about the licence and the published artefacts.

That gap is genuinely unresolved. A recital is an interpretive aid and not an operative provision, so it cannot straightforwardly add a condition that the article omits; equally, a court reading Article 53(2) purposively will read it in light of Recital 103. The practical scenario is common: weights published under a permissive licence, alongside a paid hosted API for the same model. Whether that is “monetised” in the relevant sense, and whether it matters at all given the article's silence, would be settled by AI Office guidance or a CJEU reference. Neither existed at the time of writing, and anyone asserting a clear answer is over-reading the material.

What this means for a release decision

Read as a checklist rather than as a slogan, the exemption is a decision about four things: the licence terms, what is published alongside the weights, the training compute, and the copyright and summary work that you owe regardless.

The last of those is the one that surprises teams, because it is exactly the work an open release does not automatically produce. Weights and a model card are the easy part. A defensible statement of what was crawled, how reservations were detected, and what was done about them is the same work whether the model is open or closed.