Technical

Open Source AI

AI systems or models released with open licenses allowing use, modification, and distribution of the source code and weights.

Definition

Open Source AI refers to AI systems, models, or components released under licenses that permit users to access, use, modify, and redistribute the underlying code, model weights, and/or training data. The EU AI Act provides special provisions for certain open source AI.

EU AI Act Treatment (Article 2(12)): The AI Act provides limited exceptions for open source AI:

  • Free and open-source AI components with permissive licenses may have reduced obligations
  • Exception does NOT apply to high-risk AI systems
  • Exception does NOT apply to prohibited AI practices
  • GPAI models released under open licenses still subject to Chapter V (with some modifications)

Open Source AI Levels:

  • Open weights: Model parameters publicly available (e.g., Llama, Mistral)
  • Open source: Training code and architecture available
  • Open data: Training datasets disclosed
  • Fully open: All components open and reproducible

Regulatory Considerations:

  • Liability: Open source providers may still face liability for defects
  • Downstream use: Users who deploy open source AI for high-risk purposes bear full compliance obligations
  • Modifications: Significant modifications may create new provider obligations
  • Commercial use: Using open source commercially may trigger full regulatory requirements

Related concepts: General-Purpose AI Model, Provider, AI Value Chain

Sources

  • Open Source Definition
  • EU AI Act Article 2(12)