General-purpose AI (GPAI) model

The GPAI layer is the only part of the Act aimed at the model rather than the use, which means it binds even when the provider has no idea what the model will eventually be used for. It also creates an information supply chain: a downstream company that fine-tunes or embeds a model depends on the upstream provider's Article 53 documentation to complete its own compliance file. If your vendor cannot produce that documentation, your own conformity work stalls, so it belongs in procurement questions rather than in a legal review at the end.

A SaaS company builds a customer-support assistant on a commercial foundation model. It is not a GPAI provider — it did not place the model on the market — but it is the provider of the resulting AI system, and it needs the model provider's documentation to describe capabilities and limitations in its own technical file. If instead it fine-tunes an open-weights model substantially and releases it under its own name, it may itself become a GPAI provider for the modified model.

Assuming GPAI obligations only apply to the handful of large labs. Any organisation that places a model with significant generality on the EU market under its own name is a GPAI provider, including companies distributing fine-tuned open-weight models, and open-source release does not waive the copyright-policy or training-data-summary duties.

When did GPAI obligations start applying?

2 August 2025 for models placed on the market from that date. Models already on the market before then were given a longer runway to bring documentation into line, but the substantive duties in Articles 53 to 55 are live.

Is fine-tuning a model enough to make us a GPAI provider?

It can be. If you modify a model and place the result on the EU market under your own name or trademark, you take on provider obligations for that modified model — in practice scoped to your modification, but you need your own documentation for it.

Do open-source models escape the GPAI rules?

Only partly. Free and open-source models released with public weights and parameters get relief from some documentation duties, but not from the copyright policy or the public training-data summary, and not at all if the model carries systemic risk.