The EU AI Act introduced an entirely new regulatory category that did not exist in earlier drafts: General Purpose AI (GPAI) models — also known as foundation models, large language models, or frontier AI models. Chapter V of the Act (Articles 51–56) imposes a standalone set of obligations on GPAI model providers that apply regardless of whether or how those models are used in downstream applications. If you build, fine-tune, or distribute a GPAI model in the EU market, this guide covers everything you need to know.
- GPAI obligations apply to model providers — not to the businesses that build applications on top of GPAI models. Downstream deployers have separate but lighter obligations.
- All GPAI providers must meet baseline transparency and copyright obligations. Providers of systemic risk models (above 10²⁵ FLOPs) face significantly more demanding additional requirements.
- The GPAI Code of Practice, published by the EU AI Office in 2025, provides a detailed implementation framework. Signing and following the Code creates a presumption of compliance.
- GPAI obligations are enforced directly by the EU AI Office — not national authorities — meaning global providers face a single, centrally coordinated regulator.
- What is a “General Purpose AI model” under the EU AI Act?
- Who is a GPAI model provider?
- Baseline obligations for all GPAI providers (Articles 53–54)
- Systemic risk models: the additional obligations (Article 55)
- The GPAI Code of Practice: what it covers and why it matters
- What GPAI obligations mean for downstream application builders
- The open-source exception: what it covers and its limits
- Enforcement: the EU AI Office’s direct jurisdiction
- FAQ
1. What Is a “General Purpose AI Model” Under the EU AI Act?
Article 3(63) of the EU AI Act defines a GPAI model as “an AI model, including where such an AI model is trained with 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 the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications.”
In plain language: a GPAI model is one that was trained to do many things, not one specific task. The defining characteristic is significant generality — the ability to perform competently across diverse tasks from a single underlying model. This distinguishes GPAI models from narrow task-specific AI systems (a sentiment classifier, a fraud detector, a medical image analyser) which are purpose-built for one domain.
- GPT-4, GPT-4o, o1, o3 (OpenAI)
- Claude 3 family (Anthropic)
- Gemini 1.5 Pro, Ultra (Google DeepMind)
- Llama 3, Llama 3.1 (Meta)
- Mistral Large, Mixtral (Mistral AI)
- Command R+ (Cohere)
- DALL-E 3, Sora, Stable Diffusion XL
- Fraud detection models (single task)
- Medical diagnostic AI for one condition
- Credit scoring models
- CV ranking algorithms
- Recommendation engines for one platform
- Narrow NLP classifiers (sentiment, spam)
- Speech-to-text for one language domain
2. Who Is a GPAI Model Provider?
Article 3(3) defines a provider as any person who develops a GPAI model and places it on the market. Critically for GPAI, this includes:
3. Baseline Obligations for All GPAI Providers (Articles 53–54)
Every GPAI model provider — regardless of model size, compute used, or whether it is open or closed source — must meet the baseline obligations of Article 53. These apply from 2 August 2025 (one year before the general high-risk AI deadline).
Draw up technical documentation covering: information required by Annex XI (training methodology, training data, compute used, model architecture, capabilities and limitations, evaluation results) and Annex XII (for models made available to downstream providers — summary of capabilities, limitations, and any safeguards applied).
GPAI providers must make available to downstream providers (those building applications on the model) sufficient information about the model’s capabilities, limitations, and applicable safeguards to enable those providers to fulfil their own EU AI Act compliance obligations. This creates a compliance supply chain: a downstream high-risk AI application provider cannot meet their Annex IV requirements without adequate information from their GPAI model provider.
Implement and make publicly available a policy for complying with EU copyright law — in particular the Text and Data Mining (TDM) exceptions in Articles 3 and 4 of the DSM Directive. This must address: how training data was collected, how opt-out signals were identified and respected, and what the provider’s approach is to rights holders who request removal of their content from future training runs. The CJEU’s pending ruling on LLM copyright will directly affect what this policy must say in practice.
Publish a sufficiently detailed summary of the content used for training the GPAI model — covering the types of data, sources, and geographic or temporal scope. The summary must be published in a machine-readable format. This is distinct from the full Annex XI documentation: the summary is public-facing; the full technical documentation is held internally and shared with the EU AI Office on request.
Retain technical documentation for 10 years from the date the GPAI model was placed on the market. For continuously updated models (where the model is re-trained or updated on an ongoing basis), the clock restarts from the last material update. Determine and document your version management approach — what constitutes a new model version vs. a new model for documentation purposes.
4. Systemic Risk Models: The Additional Obligations (Article 55)
A subset of GPAI models are designated as having “systemic risk” — those with the highest computational power and therefore the greatest potential for large-scale, hard-to-predict harm. These models face significantly more demanding requirements under Article 55.
Training compute > 10²⁵ floating point operations (FLOPs)
A GPAI model trained using more than 10²⁵ FLOPs is presumptively classified as having systemic risk. As of early 2026, models known to exceed or be at this threshold include: GPT-4, Gemini Ultra, Claude 3 Opus, and the largest Llama 3 variants. The Commission may update this threshold via delegated acts. Models below the threshold can also be designated systemic if the EU AI Office determines they meet systemic risk criteria through capability evaluation.
| Additional Obligation | Article | What it requires |
|---|---|---|
| Adversarial Testing & Red-Teaming | 55(1)(a) | Conduct model evaluation, including adversarial testing (“red-teaming”), to identify and mitigate systemic risks before and after release. Evaluation must cover: CBRN risk facilitation, large-scale disinformation capability, autonomous harmful action capability, and cybersecurity attack capability. |
| Incident Reporting | 55(1)(b) | Report serious incidents and possible corrective measures to the EU AI Office within defined timeframes. A serious incident for a systemic risk GPAI includes: discovered CBRN facilitation capability, large-scale cyberattack facilitation, evidence of autonomous harmful multi-step action, or significant harm to individuals arising from model outputs. |
| Cybersecurity Measures | 55(1)(c) | Implement cybersecurity protections appropriate to the systemic risk level — including protecting model weights from unauthorised access or exfiltration, securing training infrastructure, and monitoring for attempts to extract or replicate the model through adversarial querying. |
| Energy Efficiency Reporting | 55(1)(d) | Report energy consumption data to the EU AI Office where requested. The Commission is developing standardised methodologies for measuring GPAI model energy use — systemic risk providers must be able to report against these when requested. |
5. The GPAI Code of Practice: What It Covers and Why It Matters
The EU AI Office developed the GPAI Code of Practice in close consultation with major GPAI providers, researchers, and civil society throughout 2024–2025. The Code translates the high-level obligations of Articles 53–55 into practical implementation guidance with specific, measurable commitments.
Why signing matters: Under Article 56, adherence to the Code of Practice creates a presumption of compliance with the corresponding Chapter V obligations. This does not eliminate enforcement oversight — the EU AI Office retains audit powers — but it significantly reduces regulatory risk and demonstrates proactive good faith to regulators. GPAI providers that have not signed the Code face greater scrutiny and have no presumption of compliance to rely on.
- Model documentation templates aligned with Annex XI
- Training data summary format and publication standards
- Copyright compliance policy minimum requirements
- Information disclosure standards for downstream providers
- Transparency and model cards publication requirements
- Red-teaming methodology and scope requirements
- Evaluation benchmarks for systemic risk capabilities
- Model safety policy publication standards
- Incident classification and reporting procedures
- Cybersecurity baseline controls for model weights
- Energy consumption measurement methodology
6. What GPAI Obligations Mean for Downstream Application Builders
If you build applications on top of GPAI models (using GPT-4, Claude, Gemini, or Llama via API), you are not a GPAI model provider — you are a downstream application provider or deployer. You do not carry the Chapter V obligations. However, GPAI obligations affect you indirectly in three important ways:
7. The Open-Source Exception: What It Covers and Its Limits
Article 53(2) provides a significant (but not unlimited) exemption for GPAI models released under open-source licences. GPAI providers who make model weights and parameters freely available under an open-source licence may benefit from a reduced obligation set — specifically, they are exempt from the documentation requirements of Article 53(1)(a) and (b).
What open-source providers are still required to do:
8. Enforcement: The EU AI Office’s Direct Jurisdiction
Unlike high-risk AI enforcement — which sits primarily with National Market Surveillance Authorities — GPAI enforcement is centralised at the EU AI Office. This means:
