A general-purpose AI model is a model displaying significant generality, capable of competently performing a wide range of distinct tasks, and integrable into a variety of downstream systems or applications. An AI system is a machine-based system with an intended purpose that infers outputs that can influence an environment. The model provider faces Chapter V duties, while the provider of a downstream AI system faces duties that depend on that system's own classification. Neither assessment replaces the other, and a compliant model does not by itself make a downstream system compliant.
What makes a model general-purpose?
The definition in Article 3(63) turns on generality and capability across a wide range of distinct tasks, together with integrability into a variety of downstream systems or applications. Training with a large amount of data using self-supervision at scale is part of the description, but the legal test is the combination of generality, capability and integrability, not a single training method. A narrow model trained for one task, such as a single-purpose classifier for one defined input, does not display the required generality even if it uses machine learning. A model that can competently perform many distinct tasks and be reused across products is the kind of model the definition is aimed at, and its provider must be identified separately from every downstream user.
How does an AI system differ from the model inside it?
An AI system has an intended purpose, a deployment context, users, and outputs that influence an environment. A model, by contrast, is a component that can be integrated into many systems with different purposes. The same model can sit inside a recruitment tool, a customer support assistant and a document summariser, and each of those systems can carry a different classification. The system provider specifies the intended purpose, designs the surrounding workflow, and controls how outputs are presented and acted on. Those system-level choices are invisible if the analysis stops at the model. A record should therefore name the model, its version, and the system built around it, as separate lines.
For the broader regulatory framework, see EU AI Act fundamentals.
What duties follow each object?
Providers of general-purpose AI models face Chapter V duties, including technical documentation, information for downstream providers, a copyright policy, and a public summary of training content, with additional duties for models with systemic risk. Providers of AI systems face duties that depend on classification: prohibited practices screening, high-risk requirements where Article 6 is engaged, and transparency duties where Article 50 is engaged. A downstream provider cannot discharge its system duties by pointing to the model provider's documentation alone, because that documentation does not describe the downstream intended purpose, the deployment context, or the oversight design. Conversely, a model provider's Chapter V duties are not discharged by a downstream system's conformity work.
What information should flow between model and system providers?
Article 53 requires a general-purpose AI model provider to provide downstream providers with information needed to understand capabilities and limitations and to comply with their own obligations, subject to the conditions in that article. In practice, a downstream team should request the model's intended capabilities, known limitations, evaluation results relevant to its use, update and versioning practice, and any restrictions on use. The request should be recorded, and gaps should be treated as open risks, not filled by assumption. Where the model is supplied under an open-source licence, the Act provides limited exceptions to some documentation duties, but the copyright policy and training summary duties and the systemic-risk treatment still need checking. Confirm the exact exception wording before relying on it.
What happens when a model is fine-tuned or adapted?
Fine-tuning, retrieval augmentation, system prompting and workflow design can change what a downstream system does, even if the underlying model is unchanged. Those changes belong to the system analysis. A fine-tuned model may still be the same general-purpose model for Chapter V purposes, while the system built on it may acquire a new intended purpose, a new user group, or a new risk profile. The record should version the model, the adaptation, the prompts and tools that shape outputs, and the deployment configuration together. If adaptation is substantial, the Article 25 value-chain analysis must also be checked, because a change of role can follow from a change of purpose or a substantial modification.
Continue with Article 3 definitions.
What does this guide not settle?
This guide explains the model and system distinction in outline. It does not decide whether a particular model is a general-purpose AI model, whether it presents systemic risk, or whether a particular downstream product is a high-risk AI system. Those questions need the model's capability evidence, the downstream intended purpose, and the deployment facts, read against Chapters III and V in the current consolidated text. A team that cannot obtain adequate model information should record the gap, limit the deployment accordingly, and confirm the position with a competent authority or legal adviser before treating the product as assessed.
Frequently asked questions
Does a general-purpose AI model provider also become the system provider?
Not automatically. The roles are defined separately. A model provider that also builds and places a downstream system on the market under its own name can hold both roles, but the duties for each role must be assessed separately.
Can a downstream provider rely entirely on model documentation?
No. Model documentation describes the model. It does not describe the downstream system's intended purpose, deployment context, oversight design or classification, so the downstream provider still needs its own evidence and records.
Are open-source general-purpose models outside the Act?
No, not entirely. The Act provides limited exceptions for certain free and open-source models, but those exceptions do not cover every duty, and they do not apply in the same way to models with systemic risk or to downstream systems built on the model.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗European Commission regulatory framework for AI ↗