What is a general-purpose AI model (GPAI)?
A general-purpose AI model (GPAI) is, under the EU AI Act (Regulation (EU) 2024/1689), an AI model, typically trained on large amounts of data using self-supervision at scale, that displays significant generality, can competently perform a wide range of distinct tasks, and can be integrated into a variety of downstream systems or applications. Foundation models such as large language models are the paradigm case. The Act regulates the model layer separately from the systems built on it: obligations attach to the model’s provider even before anyone wraps a product around it. GPAI obligations have applied since 2 August 2025, and the Commission’s enforcement powers, including fines, have applied since 2 August 2026.
Core provider obligations (Article 53)
A provider that places a GPAI model on the EU market must keep four things in order:
- Technical documentation for the AI Office and national competent authorities (Annex XI);
- Information and documentation for downstream providers: capabilities, limitations and integration guidance (Annex XII);
- A copyright policy consistent with EU law, honouring text-and-data-mining opt-outs;
- A public training-content summary on the AI Office template;
Models with systemic risk (presumed where cumulative training compute exceeds 10^25 FLOP, Article 51) carry additional duties under Article 55: model evaluations, adversarial testing, serious-incident reporting and cybersecurity. The GPAI Code of Practice published by the Commission in July 2025 is a voluntary tool for meeting these duties. Providers of models placed on the market before 2 August 2025 have until 2 August 2027 to comply (Article 111(3)).
Models released under free and open-source licences enjoy limited exemptions from the documentation duties, unless they pose systemic risk.
The legal dimension: where the provider line sits
Under the Commission’s GPAI guidelines, modifying a model (for example by fine-tuning) with more than one third of the original training compute makes you a provider for that modification, including a training-content summary covering your fine-tuning data. Prompting and retrieval-augmented generation (RAG) generally keep you downstream. The boundary matters twice over: it decides who owes the Article 53 set, and it intersects with system-level rules when the model is integrated into a high-risk AI system, where a separate compliance regime binds the system provider.
Turkish context
Türkiye has no AI-specific statute in force; KVKK and general legislation govern domestically. The AI Act still reaches Turkish companies that place models on the EU market or fine-tune them for EU-facing products: the modification can carry provider duties of its own. EU customers increasingly pass Article 53 documentation expectations down the contract chain, so Turkish model builders tend to meet these requirements in procurement long before any regulator asks.
Do: work out whether you are a model provider, a downstream provider or a deployer before signing EU contracts, and version your technical documentation from the first release. Don’t: treat fine-tuning as “just engineering”; a substantial modification can shift Article 53 duties, including the training-content summary, onto you.
Related guides: You Didn’t Train the Model, but You Still Have AI Act Obl….
Sources. Regulation (EU) 2024/1689 (AI Act).