A general-purpose AI model, or GPAI model, is an AI model that displays significant generality and is capable of competently performing a wide range of distinct tasks, and that can be integrated into a variety of downstream systems or applications.
That test comes from Article 3(63) of the EU AI Act, which excepts AI models that are used for research, development, or prototyping activities before they are placed on the market. The EU legal term of art is the subject here, not the everyday sense of software that handles more than one job, and not artificial general intelligence.
Key Takeaways
- Article 3(63) sets a cumulative test, because a model has to display significant generality, be capable of competently performing a wide range of distinct tasks, and be capable of integration into a variety of downstream systems or applications.
- Recital 98 adds a legislative heuristic, treating models with at least a billion parameters and trained with self-supervision at scale as displaying significant generality.
- Training compute greater than 10^23 FLOP plus the ability to generate language, text-to-image or text-to-video is an indicative criterion, according to European Commission guidelines.
- Article 3(66) defines a general-purpose AI system separately, as an AI system which is based on a general-purpose AI model and which has the capability to serve a variety of purposes.
- Whoever develops a general-purpose AI model, or has one developed, and places it on the Union market under its own name or trademark is its provider, whether for payment or free of charge, per Article 3(3). Obligations for all providers of GPAI models entered into application on August 2, 2025, according to Commission guidance.
How Does the General-Purpose AI Test Work?
Three questions decide the answer, and they run in order. Taken out of order (compute before capability), they produce an argument about the wrong thing.
1. Apply the Article 3(63) Test
The statutory test asks whether a model displays significant generality and is capable of competently performing a wide range of distinct tasks. It also asks whether the model can be integrated into a variety of downstream systems or applications, regardless of the way the model is placed on the market, per Article 3(63).
Recital 98 says that models with at least a billion parameters and trained with a large amount of data using self-supervision at scale should be considered to display significant generality and to competently perform a wide range of distinctive tasks. A recital explains the legislator’s reasoning and is read alongside the Articles, so that the parameter figure works as an interpretation rather than as an operative threshold.
A driver’s license class works the same way. It is set by what the vehicle can do rather than by the badge on the hood, and the weight printed on the form sorts applications without being the rule itself.
2. Check the Commission’s Indicative Criterion
The Commission calls training compute an imperfect proxy for generality and capabilities, and still considers setting an indicative criterion which includes a training compute threshold to be the most suitable approach at present.
According to European Commission guidelines, that criterion is training compute greater than 10^23 FLOP together with the ability to generate language (whether in the form of text or audio), text-to-image or text-to-video. That threshold corresponds to the approximate amount of compute typically used to train a model with one billion parameters on a large amount of data.
A model that meets the criterion but, exceptionally, does not display significant generality or is not capable of competently performing a wide range of distinct tasks is not a general-purpose AI model. A model that does not meet the criterion but, exceptionally, displays significant generality and competently performs a wide range of distinct tasks is one.
Customs classification behaves the same way. The tariff code follows what an item is and does, and the declared weight (the quick check a border officer runs at the counter) never settles the code on its own.
3. Confirm the Model Is Placed on the Union Market
The AI Act applies to providers placing general-purpose AI models on the Union market, per Article 2(1)(a), irrespective of whether those providers are established or located within the Union or in a third country.
A provider is the natural or legal person, public authority, agency or other body that develops a general-purpose AI model or has one developed. That same actor places the model on the market under its own name or trademark, whether for payment or free of charge.
| Element of the test | Where it comes from | What it does |
|---|---|---|
| Significant generality | Article 3(63), EU AI Act | Binding condition inside the definition |
| Wide range of distinct tasks | Article 3(63), EU AI Act | Binding condition inside the definition |
| Integration into downstream systems | Article 3(63), EU AI Act | Binding condition inside the definition |
| At least a billion of parameters | Recital 98, EU AI Act | Interpretive heuristic, not an operative threshold |
| Training compute above 10^23 FLOP | Commission guidelines | Indicative criterion, and exceptions run both ways |
Sources: Official Journal of the European Union; European Commission
Why Does General-Purpose AI Matter?
The definition describes a model. The status it creates lands on an organization, and those are separate things.
The Commission considers a downstream modifier to become the provider of the modified general-purpose AI model only if the modification leads to a significant change in the model’s generality, capabilities, or systemic risk. Its indicative criterion for that point is training compute used for the modification greater than a third of the training compute of the original model.
That one-third figure is Commission guidance interpreting a gap the Act does not address, and the register matters when a compliance team quotes it back to a vendor.
Whoever holds the provider role holds the Article 53 duties: technical documentation of the model, a policy to comply with Union law on copyright and related rights, and a publicly available summary about the content used for training.
The useful property of the GPAI label is that it moves liability rather than measuring capability. A team that never pre-trained a base model can end up holding all three duties for one, and the same status question keeps recurring across instruments in our AI regulation tracking data.
GPAI Model vs GPAI System vs Foundation Model
Article 3(68) adds a further role, the downstream provider. The term means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations.
The model is what gets trained, the system is what an end user touches, and the downstream provider is whoever wires the two together. Article 53 obligations are addressed to providers of general-purpose AI models.
Foundation model is a related term from the research literature, and the Regulation does not adopt it. Presenting it as the EU equivalent of a general-purpose AI model asserts an equivalence the binding text never makes.
| Term | Where the term is defined | What it does in the Regulation |
|---|---|---|
| General-purpose AI model | Article 3(63), EU AI Act | Carries the Article 53 provider obligations |
| General-purpose AI system | Article 3(66), EU AI Act | Names the system layer built on a general-purpose AI model |
| Downstream provider | Article 3(68), EU AI Act | Names the actor integrating an AI model into an AI system |
| Foundation model | Not defined in the AI Act | No role, because the term is absent from the binding text |
Source: Official Journal of the European Union
Most of the systems people mean by general-purpose AI are large language models, and their failure modes get measured separately in our large language model hallucination data.
What Obligations Apply to Providers of GPAI Models?
Providers of general-purpose AI models must draw up and keep up-to-date the technical documentation of the model, including its training and testing process and the results of its evaluation. The purpose is to provide it, upon request, to the AI Office and the national competent authorities.
They must also put in place a policy to comply with Union law on copyright and related rights, and draw up and make publicly available a sufficiently detailed summary about the content used for training, according to a template provided by the AI Office. The documentation and summary work has a measurable price, which shows up in EU AI Act compliance cost data.
The obligations in paragraph 1, points (a) and (b), do not apply to providers of AI models released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model. The exception also requires that the parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available.
That exception does not apply to general-purpose AI models with systemic risks.
A general-purpose AI model is classified as a general-purpose AI model with systemic risk if it has high impact capabilities. Those capabilities are evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks. A model is presumed to have high impact capabilities when the cumulative amount of computation used for its training, measured in floating point operations, is greater than 10^25.
The provider then notifies the Commission without delay and in any event within two weeks after that requirement is met or it becomes known that it will be met.
Providers established in third countries must appoint an authorised representative established in the Union, by written mandate, prior to placing a general-purpose AI model on the Union market. Providers of models placed on the market before August 2, 2025 must take the necessary steps in order to comply with the obligations laid down in the Regulation by August 2, 2027.
Enforcement sits above all of it. The Commission may impose on providers of general-purpose AI models fines not exceeding 3% of their annual total worldwide turnover in the preceding financial year or €15,000,000, whichever is higher. That ceiling is discretionary, and it is a ceiling, not a fixed penalty.
| Obligation | All GPAI providers | Systemic-risk providers only | Source |
|---|---|---|---|
| Technical documentation | Yes | No | Article 53, EU AI Act |
| Copyright policy | Yes | No | Article 53, EU AI Act |
| Training-content summary | Yes | No | Article 53, EU AI Act |
| Notification to the Commission | No | Yes | Article 52, EU AI Act |
| Incident reporting | No | Yes | Commission factpage summary |
Sources: Official Journal of the European Union; European Commission
Pros, Cons, and Risks
Advantages
- Bounded supervision: The test produces a defined population to supervise instead of all AI. That is what makes enforcement tractable.
- Readable artefacts: Providers must make publicly available a sufficiently detailed summary about the content used for training, according to a template provided by the AI Office.
- A cheap initial check: The indicative criterion is one compute figure, training compute greater than 10^23 FLOP, paired with a generation capability.
Trade-offs and Risks
- Compute is an admitted proxy: The Commission calls training compute an imperfect proxy for generality and capabilities.
- The line can move: The Commission states that GPAI models are presumed to pose systemic risk if they are trained with more than 10^25 FLOP, and that this threshold is currently under review.
- The role can transfer: A fine-tuner who never trained a base model can inherit the provider duties once the modification criterion is crossed. A compute number doesn’t answer the question on its own.
- Nothing here is a safety verdict: The chapter asks for documentation, notification, assessment, and reporting. Meeting it says what a provider has recorded, not what a model will do.
Real-World Applications
Signing the Code of Practice
The General-Purpose AI Code of Practice was published on July 10, 2025. It is a voluntary tool, prepared by independent experts in a multi-stakeholder process, designed to help industry comply with the AI Act’s obligations for providers of general-purpose AI models.
Its signatories include Amazon, Anthropic, Cohere, Google, IBM, Microsoft, Mistral AI, OpenAI, ServiceNow and Aleph Alpha. The list reads as a public roll of who accepts they are in scope, and it overlaps heavily with the labs in our AI model capability comparison.
Signing Only Part of It
xAI signed up to the Safety and Security Chapter, which means it will have to demonstrate compliance with the AI Act’s obligations concerning transparency and copyright via alternative adequate means.
The obligations are the fixed thing, and the Code is one route (not the only one) to showing they have been met. Declining to sign changes the evidence a provider has to produce, not the duty itself.
Becoming a Provider by Fine-Tuning
An organization that spends modification training compute greater than a third of the training compute of the original model meets the Commission’s indicative criterion for being considered the provider of a general-purpose AI model. Article 53 duties follow, covering technical documentation, a copyright policy, and a publicly available training-content summary.
Fine-tuning is ordinary engineering practice now rather than a research exercise, a shift visible in our generative AI adoption data.
Does Fine-Tuning a Model Make You a Provider?
Not automatically. The modification has to be large enough to change what the model is, and the Commission puts a compute figure on it.
A downstream modifier that cannot be expected to know the original model’s training compute and cannot estimate it uses a replacement threshold. That threshold is a third of the threshold for a model being presumed to have high-impact capabilities, currently 10^25 FLOP, where the original model is a general-purpose AI model with systemic risk. Otherwise, it is replaced with a third of the threshold for a model being presumed to be a general-purpose AI model, currently 10^23 FLOP.
All of it is Commission guidance interpreting a gap in the Act rather than statutory text, which matters if you are weighing how much that number really carries.
Are Open-Source Models Exempt From the AI Act?
No. The exception in Article 53(2) covers paragraph 1, points (a) and (b), and it does not apply to general-purpose AI models with systemic risks.
Points (c) and (d) sit outside the carve-out (the copyright policy and the public summary of training content). The conditions are demanding too, because the weights, the architecture information and the usage information all have to be publicly available, not just a permissive license in the repository.
Conclusion
The definition is qualitative, and the operational signal is not. The Commission’s indicative criterion is training compute greater than 10^23 FLOP paired with a generation capability, an amount that corresponds to the approximate amount of compute typically used to train a model with one billion parameters on a large amount of data. Article 3(63) asks whether a model displays significant generality, whether it competently performs a wide range of distinct tasks, and whether it can be integrated into a variety of downstream systems or applications.
The Commission says the systemic-risk threshold is currently under review. Its own guidelines already treat compute as an imperfect proxy, so the number looks likelier to move than the wording does. Anyone tracking the term should watch the compute figures and expect the qualitative test to stay where it is.
