EU AI Act general-purpose AI models

A general-purpose AI model provider must keep technical documentation (EU AI Act Article 53 and Annex XI), give downstream providers what they need (Annex XII), have a copyright policy and publish a training-content summary. Models with systemic risk (Articles 51 and 55) carry more. These have applied since 2 August 2025; older models have until 2 August 2027.

The requirement in plain English

What a general-purpose model is. EU AI Act Article 3(63) defines it as a model, often trained on large amounts of data with self-supervision, that shows significant generality, can competently perform many distinct tasks and can be built into many downstream systems. Models used only for research or prototyping before release are outside it. The Commission’s guidelines of 18 July 2025 give an indicative test: training compute above 10^23 floating-point operations and the ability to generate language, text-to-image or text-to-video.

Who the provider is. The person who develops the model, or has it developed, and places it on the EU market, whether by an API, a download or by building it into their own product. Most companies using AI are not model providers: they build systems on someone else’s model, or use them.

Every provider (EU AI Act Article 53(1)). The provider:

keeps up-to-date technical documentation of the model, its training and testing, covering at least Annex XI, for the AI Office and national authorities on request;
gives providers who build on the model the information and documentation in Annex XII, so they understand its capabilities and limits and can meet their own duties;
has a policy to comply with EU copyright law, including respecting rights holders’ text-and-data-mining opt-outs;
publishes a sufficiently detailed summary of the content used for training, on the AI Office’s template.

Open release (EU AI Act Article 53(2)). A model released under a free and open-source licence, with its weights, architecture and usage information public, is exempt from the first two duties. The copyright policy and the training summary still apply, and the exemption does not cover models with systemic risk.

Systemic risk (EU AI Act Article 51 and Article 52). A model has systemic risk when it has high-impact capabilities, presumed when its cumulative training compute exceeds 10^25 floating-point operations, or when the Commission designates it on the criteria in Annex XIII. The provider notifies the Commission within two weeks of meeting the threshold, or of knowing it will, and may argue the model does not in fact present systemic risk.

Systemic-risk duties (EU AI Act Article 55). On top of Article 53, the provider evaluates the model with state of the art protocols including adversarial testing, assesses and mitigates systemic risks at EU level, reports serious incidents to the AI Office without undue delay, and protects the model and its infrastructure with adequate cybersecurity.

Providers outside the EU (EU AI Act Article 54). They appoint an authorised representative in the EU before placing a model on the market, unless the open-source exemption applies and the model has no systemic risk.

Codes and dates. The General-Purpose AI Code of Practice (EU AI Act Article 56), final on 10 July 2025, covers transparency, copyright and, for the largest models, safety and security; the Commission and the AI Board found it adequate, and following it is a way to show compliance. These duties have applied since 2 August 2025. A model placed on the market before then has until 2 August 2027 (EU AI Act Article 111(3)). The Commission can fine model providers from 2 August 2026 (EU AI Act Article 101).

This is guidance to help you understand the requirement, not legal advice.

What good looks like

A 15-person legal technology company fine-tunes an open-weight language model on contract templates and sells drafting through its own API.

Its first record is a role decision per model. For the base model, the company is a downstream provider: it keeps the original developer’s model documentation and licence, and checks them each time it updates the base version. For its fine-tuned model, it records the compute used for fine-tuning against the original model’s published training compute. The fine-tune used well under a third, so the company is not the provider of a new general-purpose model, and the memo shows the arithmetic and the sources.

It is, though, the provider of an AI system that generates text, so the Article 50 marking duty applies to its drafting output from 2 August 2026, and its high-risk screen confirms legal drafting is not an Annex III use.

The company keeps a copyright log for its fine-tuning data anyway: where each set of templates came from and on what licence. If a larger fine-tune ever crosses the one-third line, it already has the material for documentation and a training summary of its own modification.

Evidence an auditor expects

A role decision per model (usual): the AI Office or a national authority first asks whether you are the provider at all.
The compute calculation for any fine-tune: the figures and sources behind your answer to the one-third test.
Annex XI technical documentation (required for providers, unless exempt): current, and ready to hand over on request.
Annex XII information for downstream providers (required unless exempt): what you give to those who build on your model.
Copyright policy and opt-out handling (required): how rights reservations are found and respected in your training data.
Public training-content summary (required): published on the AI Office template.
For systemic-risk models: the notification, evaluation and adversarial testing results, incident reports and security measures.

Common mistakes

Treating every API customer as a model provider

Using a model through an API does not make you its provider. Your duties attach to the system you build, and getting this wrong sends effort to the wrong documents.

Assuming fine-tuning never changes your role

A heavy fine-tune can make you the provider of the modified model. Record the compute for every fine-tune, not only the large ones.

Reading the open-source exemption too widely

Open release removes the documentation duties, not the copyright policy or the training summary, and never for a model with systemic risk.

Treating the compute figures as bright lines

The 10^23 figure is an indicative criterion, and the 10^25 figure a presumption the provider may contest. Generality and capability decide; record your reasoning, not only the number.

Forgetting the 2027 deadline for older models

A model on the market before 2 August 2025 has until 2 August 2027, not forever. Plan the documentation for every model version still offered.

How Sted helps

Sted treats the EU AI Act as applying when at least one of your AI systems is in EU scope and carries obligations: it is prohibited, high-risk through the product route (it embeds a medical device under EU MDR classed above Class I) or through an Annex III use you recorded, or it is a generative system under the Article 50 transparency duties; a general-purpose model you trained and that is in EU scope counts only when it carries systemic risk; in-scope systems that are all minimal-risk give applies in part; and a stored high-risk assessment overrides the derived tier.
Which standards apply to you, and a device’s class, are worked out by Sted’s rules from your recorded answers; the AI model does not decide them.

Sted counts a model toward the Act only when you trained it. Whether a fine-tune makes you its provider is your decision, and Sted does not notify the Commission.

EU AI Act: the hub, with the AI Office’s role in enforcing the model rules.
Transparency obligations: the marking duty that applies to systems built on a model.
Risk classification: the high-risk screen for the system you build on a model.
ISO 42001 Annex A controls: the data and third-party controls that cover a model you use or supply.
Supplier due diligence guide: assessing the model provider you build on.

Frequently asked questions

Does building on a model through its API make us a general-purpose model provider?

No. Calling another company’s model through an API makes you a downstream provider or deployer of an AI system, not the provider of the model. Your duties attach to your system, such as the Article 50 marking for generated content, and you rely on the model provider’s documentation for downstream providers.

Does fine-tuning an open-weight model make us its provider?

Only if the modification is substantial. The Commission’s July 2025 guidelines use an indicative test: the compute used for the modification is more than one third of the compute used to train the original model. If you cross it, you become the provider of the modified model, with duties limited to your modification.