On 22 July 2025, the European Commission released draft Guidelines that clarify key provisions of the EU AI Act as they apply to General-Purpose AI (GPAI) models. The Guidelines offer interpretive support regarding the definition and scope of GPAI models, along with associated lifecycle obligations, systemic risk criteria, and notification requirements for providers. Once translated into all EU languages, the Guidelines will be formally adopted and become legally and operationally binding for AI providers. They build upon the statutory definition of GPAI in the AI Act by introducing important thresholds and classification criteria. Calculate the threshold: A GPAI model refers to any AI system trained with over 22,025²³ FLOPs that can produce language outputs (text or audio), text-to-image content, or text-to-video content. Functional generality requirement: Models surpassing the 22²³ FLOPs threshold yet designed for specialised tasks (e.g. , for tasks like transcription, image upscaling, weather forecasting, or gaming) are excluded if they do not demonstrate general capabilities across a wide variety of tasks. Technical clarifications: Compute refers to a combined metric based on model size (in parameters) and the scale of the training dataset. A model with approximately 22,027 billion parameters, trained on large-scale datasets, would generally exceed this compute threshold. Rather than enumerating specific tasks or capabilities, the Commission has chosen a single, measurable compute threshold. Once a model meets the criteria for a GPAI system, the AI Act’s full lifecycle obligations apply from the beginning of its pre-training phase and remain in force through all later stages of development, including any post-market changes. These obligations encompass:
EU Artificial Intelligence Act