EU Sets Guidelines for General-Purpose AI Model Providers

New Regulatory Framework for AI Models

The European Commission has issued comprehensive guidelines clarifying obligations for providers of general-purpose AI models under the EU AI Act. These regulations take effect on August 2, 2025, establishing clear compliance requirements for developers and companies working with advanced AI systems.

Key Definitions and Thresholds

The guidelines introduce technical criteria defining when an AI model qualifies as "general-purpose." Models generating text/images using over 10²² FLOPs (floating point operations) automatically fall under this category. The framework distinguishes between minor modifications and significant changes that trigger new provider obligations.

Practical Implementation Approach

The guidelines adopt a pragmatic stance:

• Only developers making substantial modifications to models face full compliance requirements
• Open-source models receive specific exemptions to encourage innovation
• Downstream modifiers become providers only when modifications exceed ⅓ of original training compute

Systemic Risk Management

Models using over 10²⁵ FLOPs are presumed to pose systemic risks. Providers of such models must conduct rigorous evaluations, implement cybersecurity measures, and report incidents to the newly established EU AI Office.

Compliance Timeline

The obligations roll out in phases:

• August 2025: Initial obligations take effect with voluntary compliance
• August 2026: Enforcement powers activate including penalty provisions
• August 2027: Full compliance required for all providers

The Commission encourages early engagement with the AI Office to navigate transitional requirements. These guidelines complement the voluntary General-Purpose AI Code of Practice, which details practical compliance methods.