We fine-tune an open-weight model for our product

Depends on the useProvider

Engineers fine-tune an open-weight model (for example Llama or Mistral) on company data and ship it inside a product.

How we classify it

The company is the provider of the AI system it ships.
It becomes the provider of a general-purpose AI model only if the modification significantly changes the model; the Commission's indicative criterion is training compute for the modification above one third of the original model's training compute. Typical fine-tuning stays far below that.
The risk class depends on what the product does: customer support is minimal or transparency; recruitment, credit or education uses are high-risk.
Result: provider of the AI system in every case; general-purpose model obligations only above the compute criterion.

Record the compute used for fine-tuning and the base model's documentation
Classify the product use with the risk procedure
Article 50 disclosure and marking if the product talks to people or generates content
Check the model licence and training data rights

Large-scale continued pre-training can cross the one-third criterion and bring general-purpose model obligations (documentation, copyright policy, training data summary)

In the AI Compliance Kit: 03AI System Inventory · 04Risk Classification Procedure · 07Vendor Due Diligence Questionnaire.

Legal basis: AI Act art. 3(3), 3(63); Commission guidelines on general-purpose AI models (18 July 2025).