In everyday terms, a GPAI model is the engine rather than the car. A foundation model that one company uses for customer support, another for code review and a third for summarizing contracts is general-purpose; each of those products is a separate AI system built on top of it. The EU AI Act regulates the two layers differently.
Providers of GPAI models must, under Article 53, keep technical documentation, give downstream builders enough information to meet their own obligations, put in place a policy to comply with EU copyright law (including machine-readable opt-outs by rights holders; see copyright and AI training), and publish a summary of the content used for training. Article 51 adds a “systemic risk” tier for the most capable models, presumed when training used more than 10^25 floating point operations, which brings extra duties such as evaluations and incident reporting. That threshold is the Act’s proxy for a frontier model.
Why it matters for builders: if you fine-tune or wrap someone else’s model, much of your compliance evidence depends on what the GPAI provider documents. Editorially, the GPAI rules are the main place the Act reaches the large model labs directly rather than through the products built on their models.
Sources
- EU AI Act Explorer: Article 3, Definitions — Article 3(63) defines a general-purpose AI model, including models trained on large amounts of data with self-supervision at scale; checked October 24, 2026.
- EU AI Act Explorer: Article 51, Classification of General-Purpose AI Models as Models with Systemic Risk — Presumes high-impact capabilities when cumulative training compute exceeds 10^25 floating point operations and lets the Commission amend thresholds; checked October 24, 2026.
- EU AI Act Explorer: Article 53, Obligations for Providers of General-Purpose AI Models — Requires technical documentation, information for downstream providers, a copyright-compliance policy and a public summary of training content; checked October 24, 2026.