General-Purpose AI Model (GPAI)

A general-purpose AI model (GPAI model) is a model trained on large amounts of data using self-supervision at scale that displays significant generality and can competently perform a wide range of distinct tasks, regardless of how it is placed on the market or integrated into downstream systems. The term is formally defined in Article 3 of Regulation (EU) 2024/1689 (the EU AI Act, point 63). Models used purely for research, development, or prototyping before any market release are excluded from the definition.

The EU AI Act definition

Article 3(63) defines a general-purpose AI model as an AI model — including one trained on a large amount of data using self-supervision at scale — that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of how it is placed on the market, and that can be integrated into a variety of downstream systems or applications. The exclusion for pre-release research and prototyping means a model still undergoing internal development, with no commercial distribution, does not yet attract GPAI obligations.

The regulation draws a deliberate line between a GPAI model and a GPAI system. A GPAI model is the model itself — the trained weights, the architecture, the capability layer. A GPAI system is a product or service built on top of that model and made available to end users. Chapter V obligations attach to the model and its provider, not to every business that deploys an application built on it. That distinction matters enormously in practice: a company integrating a foundation model into a recruitment tool is not inheriting the GPAI provider obligations. It is responsible, instead, for classifying the system it has built — which may well be high-risk under Article 6 and Annex III.

Why it matters

Chapter V of Regulation (EU) 2024/1689 establishes a dedicated compliance framework for GPAI model providers, separate from the four-tier risk classification that governs AI systems. These obligations have applied since 2 August 2025.

Baseline obligations (Article 53) apply to every GPAI model provider without exception. They are:

• Prepare and maintain technical documentation about the model (including how it was trained, its architecture, and its capabilities), using the template in Annex XI.
• Provide that documentation to downstream providers integrating the model, plus information sufficient for those providers to meet their own obligations.
• Establish and publish a policy for compliance with EU copyright law, specifically Directive 2001/29/EC and Directive 2019/790/EC.
• Publish a sufficiently detailed summary of the training content — a transparency measure aimed at copyright holders and the public, using the template in Annex XII.

These four duties form the floor. A GPAI model provider that does nothing else is still expected to satisfy all of them.

Systemic-risk obligations (Article 55) apply to a subset of GPAI models: those posing systemic risk. Article 51 defines how a model earns that designation. The primary threshold is compute: any model trained using more than 10²⁵ floating-point operations (FLOPs) is presumed to pose systemic risk. The European Commission may also designate a model as systemic-risk on the basis of other criteria — including its reach, its capability profile, or assessments provided by the scientific panel of independent experts — under the Article 51 procedure. Once designated, Article 55 requires:

• Adversarial testing (red-teaming) to identify failure modes.
• Model evaluation against the benchmarks and methodologies developed in cooperation with the AI Office.
• Tracking, documenting, and reporting serious incidents to the AI Office without undue delay.
• Adequate cybersecurity protection for the model and its infrastructure.

The supervising body is the EU AI Office, based in Brussels. For GPAI model providers — unlike the rest of the Act, where national competent authorities lead — the AI Office exercises direct oversight. It has the power to request documentation, commission independent audits, and initiate the systemic-risk designation procedure. Non-compliance by a GPAI provider can attract fines imposed by the Commission of up to €15 million or 3% of total worldwide annual turnover, whichever is higher, under Article 101.

The 2 August 2025 date is worth holding firmly in mind. The Digital Omnibus, agreed in May 2026, deferred the high-risk system obligations to 2 December 2027 (stand-alone Annex III systems) and 2 August 2028 (Annex I product safety components). That deferral does not touch Chapter V. GPAI obligations were not deferred.

Examples

The clearest examples of GPAI models are large language models (LLMs) and multimodal foundation models distributed to third parties for integration. A model trained on internet-scale text and code that can write prose, summarise documents, answer questions, generate software, and translate between languages displays "significant generality" across "a wide range of distinct tasks" — it fits the Article 3(63) definition directly.

The model-versus-downstream-system distinction becomes concrete once you trace the value chain. A model provider trains and releases the model. A software company integrates it as the reasoning layer inside a product that evaluates job applicants. That product is a downstream AI system built on a GPAI model. The product company is the AI system provider under Article 3(3) and bears the obligations that flow from the system's use — in this case, Annex III point 4 and the full high-risk stack under Articles 9–15. The model provider retains its Chapter V duties. Neither party inherits the other's obligation set.

The research and prototyping exclusion similarly becomes concrete in practice. A university research group training a large model on a public compute cluster, with no intention to release it, is not a GPAI provider. The moment any version of that model is made available to third parties — even through an API, even for free — the exclusion no longer applies.

How it affects your compliance

If you are building an application on top of a foundation model, the GPAI obligations remain with the model vendor. Your obligation is to classify the system you have built.

Start with Article 6 and Annex III. Does your system fall into one of the eight high-risk use-case areas (biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration and border control, or administration of justice)? If yes, the Article 6(3) filter may still exclude you from the high-risk tier if the system poses no significant risk of harm — but you must document that assessment. If the filter does not apply, the full high-risk requirements (Articles 9 through 15, conformity assessment under Article 43, registration under Article 49) bind you before launch.

If your system interacts directly with natural persons, Article 50 limited-risk transparency obligations apply from 2 August 2026: disclosure that the person is interacting with an AI, synthetic-content labelling, and related duties.

The GPAI obligations (Article 53, Article 55) are not your concern as a downstream builder — they attach to the model provider. You should, however, request the Article 53 documentation package from any foundation model you integrate. That package — the Annex XI technical documentation and the Annex XII training-content summary — is one input into your own Article 11 / Annex IV technical documentation file, which records the components and capabilities of the system you are placing on the market.

Confir's rule-based classification engine encodes the Article 6, Annex III, and Article 50 logic that determines where your system sits. Because the engine is deterministic — same inputs, same finding, every time — the output is audit-defensible without relying on any downstream inference layer.

Frequently Asked Questions

Is a GPAI model the same as a high-risk AI system?

No. GPAI models are a separate, cross-cutting category governed by Chapter V of the EU AI Act. The four-tier risk classification (unacceptable, high, limited, minimal) applies to AI systems classified by their intended use. A GPAI model is not automatically high-risk; it may underlie systems that are high-risk, limited-risk, or minimal-risk depending on how they are deployed. Classifying the downstream system by its use is the relevant step for most companies.

When did GPAI obligations start applying?

Chapter V obligations — Article 53 baseline duties and Article 55 systemic-risk duties — have applied since 2 August 2025. GPAI models already on the market before that date had until 2 August 2027 to bring their documentation into compliance. The Digital Omnibus deferral (high-risk deadline moved to December 2027 / August 2028) does not apply to Chapter V.

What triggers the systemic-risk designation under Article 51?

The primary trigger is a training compute threshold: any model trained above 10²⁵ FLOPs is presumed to carry systemic risk. The European Commission can also designate a model as systemic-risk based on other factors — including its capabilities, market reach, or a qualified alert from the scientific panel of independent experts under Article 90 — even if the compute threshold is not met. Providers may request a reassessment if they believe a presumption does not hold.

Do the GPAI obligations apply to open-source models?

Partially. Article 53(2) provides a specific carve-out for GPAI models released under a free and open-source licence: providers are relieved of the Annex XI technical-documentation and Annex XII downstream-information duties. However, the copyright compliance policy and the public training-content summary always survive — they are not waived by the open-source exemption. Models with systemic risk receive no open-source carve-out at all; Article 55 applies regardless of licence.

Who supervises GPAI model providers?

The EU AI Office in Brussels supervises GPAI model providers directly. This is different from the rest of the AI Act, where national competent authorities lead enforcement. The AI Office can request technical documentation, commission independent audits, initiate systemic-risk designation proceedings, and recommend that the Commission impose fines under Article 101.

What are the fines for a GPAI provider that does not comply?

Under Article 101, the European Commission may fine a GPAI model provider up to €15 million or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. These fines are separate from the Article 99 penalties that apply to providers and deployers of AI systems — the two frameworks are distinct.

Related terms

• GPAI compliance obligations — Article 53 and Article 55
• Foundation model compliance for downstream builders
• Article 51 — Systemic-risk classification of GPAI models
• Article 53 — Baseline obligations for all GPAI providers
• AI risk classification — the four tiers and how to apply them
• Glossary: AI system