Bottom line: The Commission sets a computational threshold of 10²³ FLOPs for GPAI models, while models with 10²⁵ FLOPs or higher are classified as systemic risk systems requiring comprehensive risk assessments and notification within two weeks, with providers obligated to maintain documentation, publish training data summaries, and implement security measures.
On July 18, 2025, the European Commission published draft guidelines that operationalize key provisions of the EU AI Act for general-purpose AI (GPAI) models. They define classification thresholds, govern the model lifecycle, and establish notification obligations for providers.
The new guidelines clarify the definition of general-purpose AI models through objective thresholds: a GPAI model is defined as a model trained with more than 10²³ floating-point operations and capable of generating text, audio, image, or video outputs. Specialized models such as transcription or weather forecasting systems are exempted provided they lack general capabilities across a broad spectrum of tasks. Computational capacity is understood as a combined measure of model size and training data size – a model with approximately one billion parameters would typically reach this threshold.
From the pre-training phase onwards, qualifying GPAI models are subject to comprehensive lifecycle obligations: providers must continuously update documentation, publish a summary of training data according to a regulatory template, and take copyright protection guidelines into account.
Models trained with 10²⁵ or more FLOPs are automatically classified as GPAI systems with high risk potential. They are subject to additional requirements: comprehensive risk assessments, robust cybersecurity measures, and documentation of security incidents. Providers must notify the Commission within two weeks once they reach or anticipate the 10²⁵ FLOP threshold, disclosing their calculation methods. Against a systemic risk classification, providers may appeal with robust evidence such as benchmarks or scaling laws – obligations remain in effect during the review. For the first time, reassessments may be requested six months after designation.
Source: artificialintelligenceact.eu