Chapter V
General-purpose AI models under the EU AI Act
Chapter V regulates the model, not the system built on it, so none of the risk tiers in Article 6 decide anything here. One division does: every general-purpose model carries 5 duties, and a model classified as carrying systemic risk carries 5 more.
What Chapter V governs
An AI model is not an AI system. Recital 97 draws the line: models are essential components of systems but do not constitute systems on their own, because a model needs further components — a user interface, at least — before anyone can use it. Chapter V attaches to the model layer and to whoever provides it.
That has a consequence worth stating plainly before anything else. Calling an API does not make you a provider of the model behind it. It makes you a downstream provider, with duties that attach to your system, and Article 53(1)(b) exists so that the model provider hands you what you need to meet them.
Duties on every general-purpose AI model
These bind every provider placing a general-purpose AI model on the Union market, whatever its size and whatever it is used for. Two of them can be lifted by the open-source carve-out below; two cannot.
Article 53, and Article 54 for providers outside the EU
In force since 2 August 2025. Models already on the market before that date have until 2 August 2027.
- Keep technical documentation of the model
Providers must draw up and keep up to date technical documentation of the model, covering its training and testing process and the results of its evaluation, and hand it to the AI Office or a national competent authority on request.
Models released under a free and open-source licence meeting the Article 53(2) conditions — but not if the model carries systemic risk.
Article 53(1)(a) → - Give downstream providers what they need to comply
Providers must make information and documentation available to providers of AI systems that intend to integrate the model, sufficient for them to understand its capabilities and limitations and to meet their own obligations under the Regulation.
The same Article 53(2) open-source carve-out, again excluding systemic-risk models. Intellectual property rights, confidential business information and trade secrets are preserved.
Article 53(1)(b) → - Put a copyright policy in place
Providers must have a policy to comply with Union copyright law, and in particular to identify and respect reservations of rights expressed under Article 4(3) of the Copyright in the Digital Single Market Directive, using state-of-the-art technologies.
No exemption: the Article 53(2) carve-out reaches points (a) and (b) only.
Article 53(1)(c) → - Publish a summary of the training content
Providers must draw up and make publicly available a sufficiently detailed summary of the content used to train the model, following the template provided by the AI Office.
No exemption: like the copyright policy, this sits outside the Article 53(2) carve-out.
Article 53(1)(d) → - Appoint an authorised representative in the Union
A provider established outside the Union must appoint an authorised representative established in the Union, by written mandate, before placing a general-purpose AI model on the Union market.
Providers already established in the Union.
Article 54 →
When a model carries systemic risk
There are two ways in, and only one of them involves a number. The number is widely reported as the threshold; it is a presumption, and the Commission does not need it.
High impact capabilities
Article 51(1)(a) with 51(2)
A model has high impact capabilities when they are evaluated as such on the basis of appropriate technical tools and methodologies, including indicators and benchmarks. Article 51(2) adds a presumption: a model is presumed to have them when the cumulative compute used for its training exceeds 10²⁵ floating point operations.
A decision of the Commission
Article 51(1)(b) with Annex XIII
The Commission may decide, on its own initiative or following a qualified alert from the scientific panel, that a model has capabilities or an impact equivalent to high impact capabilities, judged against the Annex XIII criteria. No compute threshold is involved in this route.
Where the Commission designates a model under the second route, it judges it against the Annex XIII criteria:
- The number of parameters of the model
- The quality or size of the data set, for example measured through tokens
- The amount of computation used for training, measured in floating point operations or indicated by estimated cost, time or energy
- The input and output modalities, and the state-of-the-art thresholds for high-impact capabilities in each
- Benchmarks and evaluations of capability, including tasks performed without additional training, adaptability, autonomy and scalability, and the tools the model can reach
- High impact on the internal market due to reach, presumed at 10 000 or more registered business users established in the Union
- The number of registered end-users
Reach alone is on that list. A model that never approaches the compute presumption can still be designated on the strength of 10 000 registered business users in the Union.
What systemic risk adds
These sit on top of Article 53, not instead of it.
- Notify the Commission within two weeks
Where a model meets the high-impact-capability condition in Article 51(1)(a), the provider must notify the Commission without delay, and in any event within two weeks of the requirement being met or of it becoming known that it will be met.
A provider may argue in the notification that the model exceptionally does not present systemic risk; until the Commission accepts that, the model is classified as carrying systemic risk.
Article 52(1) → - Evaluate the model, adversarially and on the record
Providers of systemic-risk models must perform model evaluation using standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing aimed at identifying and mitigating systemic risks.
Article 55(1)(a) → - Assess and mitigate systemic risk at Union level
Providers must assess and mitigate possible systemic risks at Union level, including their sources, arising from the development, the placing on the market or the use of the model.
Article 55(1)(b) → - Report serious incidents to the AI Office
Providers must keep track of, document and report serious incidents and possible corrective measures to the AI Office, and as appropriate to national competent authorities, without undue delay.
Article 55(1)(c) → - Protect the model and the infrastructure under it
Providers must ensure an adequate level of cybersecurity protection for the model and for the physical infrastructure of the model.
Article 55(1)(d) →
The open-source carve-out, exactly
Article 53(2) is the most misquoted provision in this chapter. It lifts two duties, not four, and only where every one of its conditions is met.
What Article 53(2) lifts
Conditions: released under a free and open-source licence allowing access, usage, modification and distribution; parameters, including the weights, made publicly available; information on the model architecture made publicly available; information on model usage made publicly available.
- Lifted
53(1)(a) technical documentation, and 53(1)(b) downstream information.
- Still owed
53(1)(c) the copyright policy, and 53(1)(d) the public training-content summary.
- Not available at all to a systemic-risk model
The exception does not apply at all to a model classified as carrying systemic risk. Recital 104 is explicit that openness is not on its own a reason to drop those obligations.
Recital 104 →
What the two annexes actually ask for
Article 53(1) points (a) and (b) each name an annex, and the two are not the same document. One goes to regulators on request; the other goes to the companies building on your model.
Annex XI — Article 53(1)(a)
For: The AI Office and national competent authorities, on request
A general description: intended tasks, the systems it can be integrated into, acceptable use policies, release date and distribution methods, architecture and parameter count, input and output modality and format, and the licence
Design specifications and training methodology, including key design choices with their rationale, and what the model is designed to optimise for
Data used for training, testing and validation: type and provenance, curation methodology, number of data points, scope, how it was obtained and selected, and the measures used to detect unsuitable sources and identifiable biases
Computational resources used to train the model, training time, and known or estimated energy consumption
Annex XII — Article 53(1)(b)
For: Downstream providers integrating the model into their systems
A general description: intended tasks, the systems it can be integrated into, acceptable use policies, release date and distribution methods, interaction with hardware or software outside the model, relevant software versions, architecture and parameter count, modality and format, and the licence
The technical means required to integrate the model into an AI system
Input and output modality and format and their maximum size, such as context window length
Information on the data used for training, testing and validation: type, provenance and curation methodology
What gets misread
Each of these is a position we have seen stated with confidence, and that the text does not support.
“Under 10²⁵ FLOP means no systemic-risk duties.”
10²⁵ is a presumption under Article 51(2), not the definition. Article 51(1)(b) lets the Commission designate a model on the Annex XIII criteria with no reference to compute at all, and reach alone — 10 000 registered business users in the Union — is one of those criteria.
Article 51(1)(b), Annex XIII →“Our model is open source, so Chapter V does not apply.”
Article 53(2) lifts points (a) and (b) only, and only where the licence and the published parameters, architecture and usage information all meet its conditions. The copyright policy and the public training-content summary still apply, and the whole carve-out falls away for a systemic-risk model.
Article 53(2) →“We fine-tuned someone else's model, so we inherit all their obligations.”
Recital 109 limits the obligations in the case of a modification or fine-tuning to that modification or fine-tuning, for example by complementing the existing technical documentation rather than recreating it.
Recital 109 →“We only call an API, so the model duties are ours too.”
Chapter V binds the provider of the model. Integrating a model into your own system makes you a downstream provider with system-level duties; Article 53(1)(b) exists precisely so that the model provider hands you what you need for them.
Articles 53(1)(b), 101 →“A code of practice is optional paperwork.”
It is the route the Act names. Until a harmonised standard is published, providers may rely on a code of practice to demonstrate compliance; a systemic-risk provider who adheres to neither must demonstrate alternative adequate means of compliance for assessment by the Commission.
Articles 53(4), 55(2), 56 →
Who enforces it
Unusually for this Regulation, not the Member State. The AI Office is the sole enforcer against providers of general-purpose AI models, which means one supervisor rather than twenty-seven — and one set of powers.
- The AI Office is the sole enforcer against providers of general-purpose AI models; national market surveillance authorities do not supervise the model layer.
- Article 101 lets the Commission fine a model provider up to 3% of its annual total worldwide turnover in the preceding financial year, or €15 000 000, whichever is higher — and it reaches a refusal to supply information or to give access for an evaluation, not only the substantive duties.
- The Digital Omnibus added Articles 75a to 75d, giving the AI Office powers to open investigations, require information by binding decision, inspect, accept binding commitments and impose periodic penalty payments.
What the Digital Omnibus changed covers the new Articles 75a to 75d in detail.
Dates
- 2 August 2025Chapter V applies — The Article 53 duties have bound providers placing a general-purpose AI model on the Union market since this date, with the Article 55 duties on top for systemic-risk models.
- 2 August 2027Models already on the market must comply — Article 111(3): a provider whose model was placed on the market before 2 August 2025 has until this date to take the steps needed to comply.
Chapter V was not deferred by the Digital Omnibus. See the full timeline for every date as amended.
Take this with you
The same material as a PDF, laid out to be read away from a screen or circulated to the people who have to act on it. Free, and it does not ask for an email.
What to do next
If you provide a model, start with the four Article 53 duties above and work out which of them your release licence actually lifts. If you build on someone else’s model, the question is a different one — run your system through the risk classifier, because your obligations come from what the system does, not from what the model is.
This page states the provisions that govern model providers. It is not legal advice, and whether a particular model meets a provision turns on facts about that model — see Article 3 for the definitions the chapter runs on.