# What does the EU AI Act require of general-purpose AI model providers?

Source: https://superintelligencelicensing.com/guides/eu-ai-act-gpai-obligations  
Publisher: ZSKFL Management  
Last updated: 2026-11-15. Not legal advice.

> EU AI Act duties for general-purpose AI model providers: documentation, copyright policy, training summary, the 10^25 systemic-risk tier, dates and fines.

## At a glance

- The EU AI Act does not license general-purpose AI models. Since August 2, 2025, every provider that places one on the EU market owes four things under Article 53: technical documentation, information for downstream developers, a copyright policy, and a public summary of training content. Qualifying open-source providers are exempt from the first two.
- A model trained with more than 10²⁵ floating-point operations is presumed to carry systemic risk. Its provider must notify the Commission within two weeks and, under Article 55, evaluate the model, mitigate systemic risks, report serious incidents and secure the model.
- The Commission’s power to fine began on August 2, 2026: up to 3% of worldwide annual turnover or €15 million, whichever is higher. Models placed on the market before August 2, 2025 have until August 2, 2027.

## What does the EU AI Act require of general-purpose AI model providers?

As of November 15, 2026, a provider of a general-purpose AI model on the EU market must document the model, pass information to the companies that build on it, keep a copyright policy, and publish a summary of what the model was trained on. Providers of qualifying open-source models without systemic risk are exempt from the first two. A provider of a model with systemic risk must also notify the Commission, test the model, manage its risks, report serious incidents and protect it against theft. None of this is a license. No one at the Commission approves a model before release.

The duties sit in Chapter V of Regulation (EU) 2024/1689, Articles 51 to 56. They attach to the model, not to what it is used for, which is what separates them from the rest of the Act. The rules for AI systems, including the high-risk regime and the transparency duties, are covered in [EU AI Act compliance](https://superintelligencecompliance.ai/guides/eu-ai-act-compliance) on SuperintelligenceCompliance.ai (related publication). This page covers model providers only.

**EU AI Act duties for general-purpose AI model providers**

| Duty | Article | Provider without systemic risk | Open-source provider without systemic risk | Systemic-risk provider, open-source or not |
|---|---|---|---|---|
| Technical documentation, kept for the AI Office on request | 53(1)(a), Annex XI | Yes | Exempt | Yes, with added content |
| Information for downstream providers | 53(1)(b), Annex XII | Yes | Exempt | Yes |
| Copyright policy | 53(1)(c) | Yes | Yes | Yes |
| Public summary of training content | 53(1)(d) | Yes | Yes | Yes |
| EU authorised representative, for providers outside the EU | 54 | Yes | Exempt | Yes |
| Notify the Commission within two weeks | 52(1) | No | No | Yes |
| Model evaluation, risk mitigation, incident reports, cybersecurity | 55(1) | No | No | Yes |

## Is my model a general-purpose AI model?

Article 3(63) defines a general-purpose AI model as one that “displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications.” Models used only for research, development or prototyping before they are placed on the market are excluded.

The definition has no number in it, so the Commission supplied one in its guidelines. Its indicative criterion is that a model is general-purpose if its training compute is greater than 10²³ floating-point operations and it can generate language, as text or audio, or text-to-image or text-to-video output. The Commission says that amount of compute is typical of a model with one billion parameters trained on a large dataset.

That line is one hundred times lower than the systemic-risk line, which means the Article 53 duties reach far more models than the handful usually called frontier. The criterion is indicative in both directions: the Commission says a model above it may exceptionally not qualify if it lacks significant generality, and a model below it may still qualify. The [compute calculator](https://superintelligencelicensing.com/compute-calculator) estimates a training run in floating-point operations, and [compute thresholds](https://superintelligencelicensing.com/guides/compute-thresholds) explains how the count is made.

## Who counts as the provider?

The provider is whoever develops a general-purpose AI model, or has one developed, and places it on the EU market, for payment or free. Where the provider is established does not matter. A U.S. or Chinese developer that makes a model available in the Union is covered, and under Article 54 a provider established outside the Union must appoint an authorised representative in the Union by written mandate before placing the model on the market.

The Commission reads “placing on the market” broadly: the first availability on the Union market through an API, a download, a cloud service or integration into an application. A model is also placed on the market when its developer builds it into its own AI system and offers that system in the Union.

The question most companies actually face is fine-tuning. The Commission’s position is that modifying someone else’s model does not usually make the modifier a provider. Its indicative criterion is that a downstream modifier becomes the provider of a new model when the compute used for the modification is greater than one third of the training compute of the original model. Below that, the modifier still owes whatever the Act requires of the AI system it builds, but not the Chapter V model duties. Where the modifier does become a provider, recital 109 limits its Article 53 duties to the modification.

## What must every provider do?

Article 53(1) lists four duties.

- **Technical documentation.** The provider must “draw up and keep up-to-date the technical documentation of the model, including its training and testing process and the results of its evaluation,” containing at least the Annex XI items, “for the purpose of providing it, upon request, to the AI Office and the national competent authorities.” Annex XI asks for, among other things, the architecture and number of parameters, the acceptable use policies, the type and provenance of training data, the computational resources used to train the model, and its known or estimated energy consumption.
- **Information for downstream providers.** The provider must give companies that integrate the model enough to “have a good understanding of the capabilities and limitations” of the model and to meet their own obligations, with at least the Annex XII items. Trade secrets and confidential business information remain protected.
- **Copyright policy.** The provider must “put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790.” That reservation is the text-and-data-mining opt-out a rightsholder can declare.
- **Training content summary.** The provider must “draw up and make publicly available a sufficiently detailed summary about the content used for training,” using the template the AI Office provides. The Commission published the template and an explanatory notice on July 24, 2025.
The first duty is not a filing. The documentation is kept and produced when the AI Office or a national authority asks for it. Only the training summary is public, and only the downstream information is handed to customers. Article 53(3) adds a general duty to cooperate with the Commission and national authorities. How the training-data duty interacts with licensing deals and litigation is covered in [training data licensing](https://superintelligencelicensing.com/guides/ai-training-data-licensing).

## What does the open-source exception cover?

Less than its name suggests. Under Article 53(2), the documentation duty and the downstream-information duty “shall not apply to providers of AI models that are 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, the information on the model architecture, and the information on model usage, are made publicly available.” Article 54(6) lifts the authorised-representative duty on the same terms.

Three limits apply:

- The copyright policy and the public training summary still apply to open-source providers.
- “This exception shall not apply to general-purpose AI models with systemic risks.” An open-weight model above the systemic-risk line owes everything.
- The Commission’s guidelines read the licence condition strictly. No payment or other monetisation may be required for access, use, modification or distribution, and the guidelines list limits to non-commercial or research-only use, bans on distribution, and extra licensing above a user threshold as restrictions that disqualify a licence.
On the Commission’s reading, a licence with such terms falls outside the exception even though the weights can be downloaded. The guidelines are not legally binding; an authoritative interpretation can only come from the Court of Justice of the European Union. Which model licences carry these restrictions is compared in [open-weight model licenses](https://superintelligencelicensing.com/guides/open-weight-model-licenses).

## When does a model have systemic risk?

A model gets there by one of two routes. Under Article 51(2), a model “shall be presumed to have high impact capabilities” when “the cumulative amount of computation used for its training measured in floating point operations is greater than” 10²⁵. Under Article 51(1)(b), the Commission can also designate a model, on its own initiative or after a qualified alert from the scientific panel, using the Annex XIII criteria: parameters, dataset size, training compute, modalities, benchmark results, reach and registered end-users. Annex XIII presumes high impact on the internal market when a model “has been made available to at least 10 000 registered business users established in the Union.”

Article 52 sets out what happens next:

- **Notify.** The provider “shall notify the Commission without delay and in any event within two weeks after that requirement is met or it becomes known that it will be met.”
- **Contest.** With the notification, the provider may present “sufficiently substantiated arguments” that the model, exceptionally, does not present systemic risks because of its specific characteristics. If the Commission rejects them, the model is treated as having systemic risk.
- **Review.** A provider whose model the Commission designated can request reassessment, at the earliest six months after the decision, with “objective, detailed and new reasons.”
- **Publication.** The Commission must ensure that a list of models with systemic risk is published and kept up to date.
The words “or it becomes known that it will be met” put the clock ahead of the training run. Recital 112 explains, and the Commission’s guidelines repeat, that training takes upfront allocation of compute, so a provider can know before training finishes that a model will cross the line, and the two weeks start then. The Commission can change the threshold by delegated act (Article 51(3)). We found no delegated act changing it as of November 15, 2026.

## What do providers of systemic-risk models owe?

Everything in Articles 53 and 54, plus four duties in Article 55(1). They must:

- “perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model”;
- “assess and mitigate possible systemic risks at Union level, including their sources, that may stem from the development, the placing on the market, or the use” of the model;
- “keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them”;
- “ensure an adequate level of cybersecurity protection” for the model and its physical infrastructure.
Their technical documentation must also describe, among other things, their evaluation strategies and results and, where applicable, their adversarial testing (Annex XI, Section 2).

The Act’s incident duty has no number of hours or days in it. It says “without undue delay.” The Code of Practice does set deadlines for the providers that sign its Safety and Security chapter. Under Measure 9.3, save in exceptional circumstances, an initial report goes to the AI Office within two days of the provider becoming aware that its model was involved in a serious and irreversible disruption of critical infrastructure, five days for a serious cybersecurity breach, including exfiltration of model weights, 10 days for a death, and 15 days for serious harm to health, property or the environment or an infringement of Union law protecting fundamental rights. California’s SB 53 gives 15 days to report to the state and New York’s RAISE Act gives 72 hours. Both also require a developer to disclose an incident that poses an imminent risk of death or serious physical injury within 24 hours to an appropriate authority, such as a law enforcement or public safety agency, and neither says whether that disclosure replaces the regular report or adds to it. The U.S. state laws are set out in [California SB 53](https://superintelligencelicensing.com/guides/california-sb-53-requirements) and [New York RAISE Act requirements](https://superintelligencelicensing.com/guides/new-york-raise-act-requirements).

## What is the Code of Practice, and does a provider have to sign it?

No. The General-Purpose AI Code of Practice is voluntary. It was published on July 10, 2025 and has three chapters: Transparency, Copyright, and Safety and Security. The Commission says the Transparency and Copyright chapters offer all providers a way to demonstrate compliance with Article 53, and the Safety and Security chapter is relevant only to providers of models with systemic risk under Article 55. In an opinion of August 1, 2025, the Commission concluded that the Code adequately covers the obligations in Articles 53 and 55, and it says the AI Board confirmed that assessment.

Signing has a legal effect, but a limited one. Under Articles 53(4) and 55(2), providers “may rely on codes of practice” to demonstrate compliance “until a harmonised standard is published.” The Commission says adherence is a means to demonstrate compliance “while not providing a presumption of conformity.” A provider that does not sign “shall demonstrate alternative adequate means of compliance for assessment by the Commission,” and the Commission says such a provider must be able to justify its approach, for instance with a gap analysis against the Code. When it sets a fine, the Commission must also take into account commitments made in a code of practice (Article 101(1)).

As of the Commission’s signatory page, last updated October 7, 2026, the listed signatories include Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. The page says xAI signed only the Safety and Security chapter, so it must show compliance on transparency and copyright by other means. Meta said in July 2025 that it would not sign.

The 2026 omnibus, Regulation (EU) 2026/1744, changed how a code is endorsed. It replaced Article 56(6), which had allowed the Commission to approve a code by implementing act and give it general validity in the Union. The new text has the Commission assess whether a code covers the obligations and publish that assessment, with no approval step. The duties themselves, and their dates, were left as they were.

## How is it enforced, and how large are the fines?

The Commission enforces Chapter V itself, through the AI Office. Article 88 gives it “exclusive powers to supervise and enforce” the model duties, so national authorities do not enforce them. Article 75(1) separately gives the AI Office supervision of AI systems built on a general-purpose model by the same provider. The 2026 omnibus rewrote that paragraph: it makes the competence exclusive, extends it to providers that form part of the same undertaking and to AI systems in very large online platforms and search engines, and carves out exceptions, including AI in products covered by Annex I.

The AI Office’s powers are set out in Articles 91 to 93:

- request the documentation a provider drew up under Articles 53 and 55, or any additional information needed to assess compliance;
- evaluate the model itself, with access “through APIs or further appropriate technical means and tools, including source code”;
- request that a provider take measures to comply, implement mitigations, or “restrict the making available on the market, withdraw or recall the model.”
That last power is the closest the Act comes to a license, and it runs the other way: the Commission cannot refuse a model before release, but it can request that one be pulled afterwards, and a provider that intentionally or negligently fails to comply with a requested measure can be fined under Article 101(1)(c).

Under Article 101, the Commission may fine a provider up to 3% of its annual total worldwide turnover in the preceding financial year or €15 million, whichever is higher, where it finds the provider intentionally or negligently infringed the Regulation, failed to answer an information request or answered it incorrectly, ignored a requested measure, or refused access for an evaluation. Article 113(b) held the fining article back from the August 2025 start, and the Commission states that its enforcement powers entered into application on August 2, 2026.

Enforcement has started with questions, not penalties. On October 14, 2026, Euronews reported that the AI Office had sent requests for information to OpenAI and Anthropic, after Henna Virkkunen, the Commission’s Executive Vice-President for Tech Sovereignty, Security and Democracy, said the Office had started enforcing the rules. Legal experts quoted in the report disputed whether the Act reaches incidents that occurred during testing, before the models were placed on the EU market. A request for information is not a finding of a breach. We found no fine imposed under Article 101 as of November 15, 2026.

## What are the key dates?

**Dates for general-purpose AI model providers under the EU AI Act**

| Date | What happened or happens |
|---|---|
| July 10, 2025 | General-Purpose AI Code of Practice published |
| July 18, 2025 | Commission approves the content of its guidelines on the scope of the obligations; formally adopted November 19, 2025 |
| July 24, 2025 | Template for the public summary of training content published |
| August 2, 2025 | Chapter V obligations apply to models placed on the market from this date |
| July 27, 2026 | Omnibus Regulation (EU) 2026/1744 enters into force; model duties and dates unchanged |
| August 2, 2026 | Commission enforcement powers, including fines, apply |
| August 2, 2027 | Deadline for models placed on the market before August 2, 2025 |

The 2027 date is a transition for models that were already on the market. Article 111(3) says providers of models “placed on the market before 2 August 2025 shall take the necessary steps in order to comply with the obligations laid down in this Regulation by 2 August 2027.” A new model released today has no grace period.

## How does this compare with a license, and with U.S. rules?

The EU regime is notification and documentation backed by an after-the-fact recall power. It has no application, no approval and no permit. Its systemic-risk threshold, 10²⁵ operations, is ten times lower than the 10²⁶ line in California and New York, so it reaches models those laws do not. It also has a tier they lack: the Article 53 duties for every general-purpose model, including small and open-weight ones.

The side-by-side comparison of the EU, California, New York, China and South Korea is in [frontier AI laws](https://superintelligencelicensing.com/guides/frontier-ai-laws), and the difference between a filing, a notice and a permit is in [AI licensing explained](https://superintelligencelicensing.com/guides/ai-licensing-explained). Dates and figures here are stated as of November 15, 2026. This guide is general information, not legal advice.

## Sources

- [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj), Official Journal of the European Union, July 12, 2024: Articles 3(63), 51 to 56, 75, 88 to 94, 101, 111 and 113, Annexes XI to XIII, recitals 97 to 112.
- [Article 51: Classification of general-purpose AI models as general-purpose AI models with systemic risk](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-51), AI Act Service Desk, European Commission.
- [Article 52: Procedure](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-52), AI Act Service Desk, European Commission.
- [Article 53: Obligations for providers of general-purpose AI models](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53), AI Act Service Desk, European Commission.
- [Article 54: Authorised representatives of providers of general-purpose AI models](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-54), AI Act Service Desk, European Commission.
- [Article 55: Obligations of providers of general-purpose AI models with systemic risk](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-55), AI Act Service Desk, European Commission.
- [Article 101: Fines for providers of general-purpose AI models](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-101), AI Act Service Desk, European Commission.
- [Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/eli/reg/2026/1744/oj?locale=en), EUR-Lex, July 24, 2026: amendments to Articles 56(6) and 75.
- [Guidelines for providers of general-purpose AI models](https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers), European Commission, last updated April 28, 2026.
- [Guidelines on the scope of the obligations for providers of general-purpose AI models under the AI Act](https://digital-strategy.ec.europa.eu/en/library/guidelines-scope-obligations-providers-general-purpose-ai-models-under-ai-act), European Commission: content approved July 18, 2025 (C(2025) 5045 final), formally adopted November 19, 2025 (C(2025) 7719 final).
- [Q&A: Guidelines on obligations for general-purpose AI providers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers), European Commission.
- [General-Purpose AI Models in the AI Act: Questions & Answers](https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers), European Commission, last updated September 9, 2025.
- [The General-Purpose AI Code of Practice](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai), European Commission, last updated October 7, 2026 (chapters and signatories).
- [Code of Practice for General-Purpose AI Models: Safety and Security Chapter](https://ec.europa.eu/newsroom/dae/redirection/document/118119), European Commission, published July 10, 2025: Commitment 9, Measure 9.3 (reporting timelines).
- [Commission Opinion on the assessment of the General-Purpose AI Code of Practice](https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-general-purpose-ai-code-practice), European Commission, August 1, 2025.
- [Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, July 24, 2025.
- [EU’s AI rules face first enforcement test over lab incidents](https://www.euronews.com/2026/10/14/eus-ai-rules-face-first-enforcement-test-over-lab-incidents), Euronews, October 14, 2026.
- [Meta says it won’t sign Europe AI agreement, calling it an overreach that will stunt growth](https://www.nbcsandiego.com/news/business/money-report/meta-says-it-wont-sign-europe-ai-agreement-calling-it-an-overreach-that-will-stunt-growth/3871383/), CNBC via NBC San Diego, July 18, 2025.

## Common questions

**Do general-purpose AI providers need EU approval before releasing a model?**

No. The EU AI Act has no license or pre-release approval for general-purpose AI models. Providers must document the model, inform downstream developers, keep a copyright policy and publish a training content summary. Providers of qualifying open-source models without systemic risk are exempt from the first two. Providers of models presumed to carry systemic risk, those trained with more than 10^25 floating-point operations, must notify the Commission within two weeks. The Commission can later request that a model be restricted, withdrawn or recalled, and it can fine a provider that intentionally or negligently fails to comply.

**Is the General-Purpose AI Code of Practice mandatory?**

No. The Code, published July 10, 2025, is voluntary. Under Articles 53(4) and 55(2), providers may rely on it to demonstrate compliance until a harmonised standard is published. A provider that does not sign must demonstrate alternative adequate means of compliance for assessment by the Commission. The Commission says adherence does not give a presumption of conformity.

**When do the EU rules apply to general-purpose AI models released before August 2025?**

By August 2, 2027. Article 111(3) gives providers of models placed on the market before August 2, 2025 until that date to comply. Models placed on the market on or after August 2, 2025 had to comply from the start, and the Commission’s power to fine, up to 3% of worldwide annual turnover or 15 million euros, whichever is higher, has applied since August 2, 2026.

