Foundation Model
A foundation model is a large AI model pre-trained on vast, mostly unstructured amounts of data that serves as the basis for numerous downstream applications, such as chatbots, image generators or search assistants. Well-known examples are the GPT, Claude, Gemini and Llama models. Fine-tuning or targeted prompting adapts a single foundation model to many different tasks without retraining it from scratch.
In practice
What matters for companies is that training your own foundation model is extremely compute- and cost-intensive and in practice remains the preserve of large AI labs. Businesses instead use existing foundation models through APIs or chat interfaces and adapt them to their own content with prompting, RAG or fine-tuning. For some of these models the EU AI Act uses the separate legal term "general-purpose AI model" (GPAI, Article 3(63)): it requires significant generality and the ability to perform a wide range of distinct tasks, so it does not automatically cover every foundation model. Separate transparency and documentation obligations have applied to GPAI models since 2 August 2025; for models placed on the market before that date, Article 111(3) provides a transitional period until 2 August 2027.