What Are General-Purpose AI Models (GPAI)?
Under the EU AI Act (Regulation 2024/1689), a general-purpose AI model (GPAI) is defined as an AI model that demonstrates significant generality and can competently perform a wide range of distinct tasks, regardless of how it is placed on the market. These models are capable of integration into various downstream AI systems or applications. Models developed solely for research, development, or prototyping are excluded from this definition.
Indicative Criteria for GPAI Models
The European Commission outlines two main indicators:
- Training Compute Threshold:
Models with training compute exceeding 10^23 FLOPs are considered likely to qualify as GPAI. This usually corresponds to models trained with ~1 billion parameters and large datasets. - Multimodal Capabilities:
The model must be capable of generating:- Text (including code),
- Audio (including speech),
- Text-to-image, or
- Text-to-video content.
Even if a model meets the compute threshold, it won’t be classified as GPAI if it lacks generality. Conversely, models that don’t meet the threshold but exhibit significant general-purpose abilities may still fall under the GPAI classification.
GPAI Model Lifecycle Obligations
The lifecycle of a GPAI model begins at initial pre-training and includes all fine-tuning or post-training modifications. During the entire lifecycle, providers must:
- Maintain updated technical documentation,
- Publish a public summary of training data,
- Define a copyright policy,
- Assess and mitigate systemic risks.
GPAI Models with Systemic Risk
A GPAI model is considered high-risk if it:
- Demonstrates high-impact capabilities (e.g. comparable to leading frontier models), and
- Has training compute exceeding 10^25 FLOPs.
In such cases:
- Providers must notify the European Commission within two weeks of reaching the threshold or becoming aware that it will be met.
- The Commission may classify the model as systemic-risk GPAI based on criteria in Annex XIII of the AI Act.
- Providers may challenge this classification by presenting evidence that the model does not pose systemic risks.
Who Is Considered a Provider?
A provider is any natural or legal person who:
- Develops and places a GPAI model on the EU market under their name or trademark.
- Distributes the model through APIs, software libraries, cloud platforms, apps, or on-premise installations.
- Uses the model in internal processes that affect EU residents’ rights.
Even downstream actors who modify or deploy GPAI models may be classified as providers — especially if their modifications significantly affect the model’s generality, capabilities, or risk level.
Open-Source GPAI Models: What Are the Exemptions?
Open-source GPAI models are exempt from certain obligations if they meet all of the following:
- Released under a license allowing access, use, modification, and redistribution;
- Available free of charge (no monetization, including data monetization or gated access);
- Publicly share parameters, architecture, and usage details.
Exemptions DO NOT apply to:
- Models with systemic risk,
- Obligations related to copyright policy and training data transparency.
Enforcement, Compliance, and Penalties
- Effective Date: Obligations for GPAI providers begin on 2 August 2025.
- Supervision Authority: The AI Office (under the European Commission) handles enforcement.
- Codes of Practice may help demonstrate compliance and reduce penalties.
- Fines may reach up to 3% of global annual turnover or €15 million, applicable from 2 August 2026.
Transitional Provisions
- Models placed on the market before 2 August 2025 must comply by 2 August 2027.
- Models released after 2 August 2025 must comply immediately.
- During the first year (until 2 August 2026), the Commission will not impose fines, but compliance is still mandatory.
Metaphor: Navigating AI Compliance in the EU
Think of these guidelines as a map for autonomous vehicles. The EU AI Act defines the rules of the road, while the AI Office acts as the traffic authority, offering signposts, guidance, and transitional periods. Providers must ensure their AI “vehicles” (models) follow the route safely, transparently, and in compliance with EU law.
Stay Compliant – Stay Ahead
If you’re developing or deploying general-purpose AI models in or for the EU, it’s time to:
- Evaluate your training compute and capabilities,
- Identify if your model may pose systemic risks,
- Prepare the necessary documentation and procedures,
- Engage early with the AI Office.