Model Weights

Model Weights (also called model parameters) are the learned numerical values in a trained AI model that determine how it processes inputs and generates outputs. They represent the "knowledge" acquired during training and are central to discussions of AI intellectual property and trade secrets.

Technical Definition:

• In neural networks: connection strengths between neurons
• Learned through training on data
• Determine model behavior for any given input
• Stored as matrices of numerical values
• Large models may have billions of parameters

Legal Status:

• Trade secrets: Weights often protected as confidential business information
• Copyright: Uncertain—weights are numerical data but encode learned patterns
• Patents: Training methods may be patentable, but weights themselves typically not
• Open weights: Some models release weights publicly (Llama, Mistral)

EU AI Act Relevance:

• Technical documentation: Description of model architecture and training required
• GPAI transparency: Information about model capabilities, not necessarily weights
• Open source exception: Models with publicly available weights may have reduced obligations

Security Considerations:

• Weights can be reverse-engineered or stolen
• Model extraction attacks can approximate weights
• Protecting weights is key to AI competitive advantage

Related concepts: Foundation Model, Open Source AI, Technical Documentation