Glossary
General-purpose AI model
Written with AI from the sources below and checked against our facts sheet. Last checked . How this site is made · Spot an error?
Part of: EU AI Act
A general-purpose AI model is an AI model, such as a large language model, that can competently perform a wide range of distinct tasks. It shows significant generality and can be built into many downstream systems. Under the EU AI Act (Art. 3(63)), its provider has its own obligations, which have applied since 2 August 2025.
In plain English
A general-purpose AI model is like an engine sold to many carmakers. The engine maker doesn't know whether it will power a taxi, an ambulance or a race car. So the EU AI Act regulates the engine maker for what it alone controls: documentation, copyright and training data. The most powerful engines also need safety testing. Each carmaker is regulated for the vehicle it builds.
That split is the key idea. A GPAI model carries model duties. A GPAI system built on it, such as a chatbot app, is classified by its use like any other AI system.
Example
Northlake AI, a fictional company based in Canada, trains a large language model with about 4 × 10²⁴ FLOP of compute. It offers the model to EU businesses through an API.
- The model generates text and handles many distinct tasks, so it is a GPAI model. Commission guidelines treat more than 10²³ FLOP plus the ability to generate language, images or video as an indicative sign.
- It is below 10²⁵ FLOP, so systemic risk is not presumed. Northlake owes the Art. 53 duties and, as a non-EU provider, must appoint an authorized representative in the EU (Art. 54).
Pellet Hire, a fictional EU recruitment firm, builds a candidate-ranking tool on the model and sells it under its own brand. Pellet Hire is the provider of a high-risk AI system (Annex III, employment). Northlake's model does not become high-risk because of that use.
If Northlake later trains a successor with 2 × 10²⁵ FLOP, systemic risk is presumed. It must notify the Commission within two weeks (Art. 52) and take on the Art. 55 duties.
Why it matters
Chapter V of the EU AI Act (Arts. 51 to 56) sets two tiers:
| Who | Main duties |
|---|---|
| Every GPAI provider (Art. 53) | Technical documentation for the AI Office; information for downstream providers; an EU copyright policy that respects text-and-data-mining opt-outs; a public training-content summary |
| Open-source GPAI (free license, public weights) | Exempt from the two documentation duties, but not from the copyright policy or training summary |
| Systemic-risk GPAI (Art. 55) | All of the above, plus model evaluation with adversarial testing, assessing and mitigating systemic risks, reporting serious incidents to the AI Office, and cybersecurity |
Systemic risk is presumed above 10²⁵ FLOP of training compute (Art. 51(2)). The voluntary GPAI Code of Practice (10 July 2025) has Transparency, Copyright, and Safety and Security chapters. Providers can rely on it to demonstrate compliance until harmonized standards exist (Arts. 53(4) and 55(2)).
The obligations have applied since 2 Aug 2025. The Digital Omnibus deferred high-risk dates but left the GPAI rules untouched.
How the AIGP exam tests it
The BoK gives GPAI models their own indicator (lesson II.C.8).
Common traps:
- Model vs system. A customer using a GPAI model for hiring does not make the model high-risk. The hiring system is high-risk.
- The wrong number. 10²⁵ FLOP is the EU systemic-risk presumption. California's SB 53 uses 10²⁶ operations for frontier models. Crossing 10²⁵ adds duties; it is not a ban.
- Open-source overreach. The exemption never covers systemic-risk models, the copyright policy or the training summary.
- High-risk tools applied to models. CE marking and the EU high-risk database are for high-risk systems, not GPAI models.
The tell: first ask whether the question is about the model or the product built on it.
Frequently asked questions
What is the difference between a GPAI model and a GPAI system?
The model is the underlying engine, such as a large language model. A GPAI system is an AI system built on that model, such as a chatbot app (Art. 3(66)). The model's provider has Chapter V duties. The system is classified by its use like any other AI system, so a GPAI system used to screen job applicants is high-risk.
When does a GPAI model have systemic risk?
When it has high-impact capabilities. These are presumed when cumulative training compute exceeds 10²⁵ floating-point operations (Art. 51(2)). The Commission can also designate a model using the Annex XIII criteria, such as reach and number of users. Providers must notify the Commission within two weeks of meeting the threshold and can argue their model is an exception.
Do open-source GPAI models have obligations under the EU AI Act?
Yes, but fewer. Models released under a free and open-source license with public weights are exempt from the two documentation duties (Art. 53(2)). They still need a copyright policy and a public training-content summary. The exemption does not apply at all to models with systemic risk.
When did the GPAI obligations start?
On 2 August 2025. Models already on the market before that date have until 2 August 2027 to comply (Art. 111(3)). The AI Office can fine GPAI providers from 2 August 2026, up to €15 million or 3% of worldwide turnover, whichever is higher. The 2026 Digital Omnibus did not defer these rules.
Related
Sources
- Regulation (EU) 2024/1689 (EU AI Act), Arts. 3, 51 to 56, 101 and 111, EUR-Lex
- EU AI Act Service Desk, Article 51: Classification of GPAI models with systemic risk
- EU AI Act Service Desk, Article 53: Obligations for providers of GPAI models
- European Commission, Guidelines for providers of general-purpose AI models
- European Commission, The General-Purpose AI Code of Practice
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex