The European Commission published yesterday a 90-page guidelines document clarifying the concrete application of the AI Act for general-purpose AI models (GPAI). The text specifies in particular the compute thresholds that trigger enhanced obligations, transparency requirements toward the AI Office, and mandatory red-teaming protocols.
Key points include: models trained with more than 10^25 FLOP are subject to systemic obligations, which include risk assessment, serious incident reporting, and technical documentation accessible to authorities. The guidelines also specify that open-source models benefit from partial exemptions under certain conditions.
Affected companies have 12 months to comply. Several players, including Mistral and Hugging Face, welcomed the clarification while pointing to persistent gray areas around the precise definition of the required red-teaming.