Module 4 — GPAI Models: Who Is the Provider After You Fine-Tune?

Module 4 of the free EU AI Act Fundamentals course: general-purpose AI obligations under Art. 53-55, the systemic-risk threshold, and the one-third compute rule that decides whether fine-tuning makes you a provider.

General-purpose AI is regulated on its own track. It is not a risk tier — a GPAI model is not "high-risk" or "minimal risk"; it carries model-level obligations that sit alongside the system-level regime you met in Modules 2 and 3.

The obligations have been in force since 2 August 2025.

What counts as a GPAI model

Art. 3(63): a model displaying significant generality, capable of competently performing a wide range of distinct tasks, and that can be integrated into a variety of downstream systems. Trained on large amounts of data, typically with self-supervision at scale.

The distinction that matters downstream: a model is not a system. The model is the weights and what they can do. A system is the model plus the interface, the guardrails and the purpose. You can be the provider of a system built on someone else's model — that is the normal case, and it is exactly what happens when you build an assistant on a frontier API.

What a GPAI provider owes (Art. 53)

• Technical documentation of the model: training and testing process, evaluation results.
• Information to downstream providers so they can understand capabilities and limitations and meet their own duties.
• A copyright policy, including respecting text-and-data-mining reservations under the DSM Directive.
• A sufficiently detailed public summary of the training content, following the AI Office template.

Open-source models released under a free licence with public weights and parameters are exempt from the documentation and downstream-information duties — but not from the copyright policy or the training-content summary, and not at all if the model has systemic risk.

Systemic risk (Art. 51 and 55)

A model is presumed to carry systemic risk when the cumulative compute used for training exceeds 10²⁵ FLOPs. The Commission can also designate a model on other criteria.

Above that line, Art. 55 adds:

• model evaluation, including adversarial testing;
• assessment and mitigation of systemic risks at Union level;
• tracking, documenting and reporting serious incidents to the AI Office;
• adequate cybersecurity for the model and its physical infrastructure.

For context: no enterprise fine-tune approaches 10²⁵ FLOPs. This tier is about a handful of frontier models.

The question that actually comes up: does fine-tuning make me a provider?

This is where organisations most often over-classify themselves, and the answer is usually no.

The Commission guidelines of 18 July 2025 set an indicative criterion: you become the provider of the modified model when the compute used for the modification is greater than about one third of the compute used to train the original model. Where the original figure is not known, the guidelines point to one third of 10²³ FLOPs as the reference.

Put that in perspective. A LoRA adapter or an instruction fine-tune run by a bank on its own corpus is typically several orders of magnitude below that threshold. You are not the provider of the model. The upstream provider remains the provider, with its Art. 53 duties intact.

What you are, in that case, is the provider of the AI system you built around it — and that is where your obligations live. If that system does consumer credit scoring, it is high-risk under Annex III 5(b) and everything in Module 3 applies. If it answers customer questions, it is caught by Art. 50 transparency. The model's regime and your system's regime are separate questions with separate answers.

Where the two tracks meet

A single deployment can touch all five rows. They do not collapse into one obligation set, and treating them as one is how programmes end up either massively over-scoped or missing the one duty that actually bit.

Practical consequence for a deployer

If you consume GPAI through an API:

1. Get the Art. 53 information from your provider and keep it. You will need it for your own technical documentation if your system is high-risk — Annex IV §2(d) asks about the data.
2. Do not assume the provider's compliance is yours. Their Art. 53 documentation covers their model, not your system.
3. Track your fine-tuning compute if you fine-tune at all. Not because you are likely to cross the line, but because being able to show you are three orders of magnitude below it ends the conversation in one slide.
4. Watch for the change of purpose. Using a general model for an Annex III use case makes you the provider of a high-risk system, however small your fine-tune was. That is Art. 25, not Art. 53 — a different trap, covered in Module 3.

Check yourself

1. You fine-tune an open-weights model with a LoRA on 50,000 internal documents. Are you the model provider? — No. Nowhere near one third of the original training compute. You are the provider of the system you built.
2. Your GPAI vendor says "we are AI Act compliant". Are you covered? — No. Their Art. 53 duties are model-level. Your system's classification and obligations are yours.
3. What is the systemic-risk presumption threshold? — 10²⁵ FLOPs cumulative training compute (Art. 51).
4. When did GPAI obligations start applying? — 2 August 2025. Unlike the Annex III high-risk regime, they were not deferred.

Previous: Module 3 — Obligations by role · Next: Module 5 — Deadlines and enforcement →