In the past, a separate model was developed for every AI task: one for translation, one for text recognition, one for sentiment analysis. Foundation models have changed this approach. A single, very large model is trained once at great effort and then serves as the basis for numerous applications. The term was coined in 2021 by researchers at Stanford University.
For companies, this means: today, AI can usually be used without training your own model. The central decision is which foundation model to use and how to adapt it.
How does a foundation model work?
Foundation models are pre-trained with large amounts of general data, often using self-supervised learning. The model learns without hand-labeled data, for example by predicting missing or subsequent parts of a text. This creates a broad understanding of language, images or other data that can be transferred to many tasks.
The best-known examples are large language models. There are also foundation models for images, audio, program code or several data types at once, see multimodal AI.
A foundation model is adapted in three ways, often in combination:
- through instructions and examples in the prompt
- by connecting your own documents at runtime
- by further training with your own data, that is, fine-tuning
Typical examples
- A language model is integrated into a ticketing system via an application programming interface.
- An open model runs on your own hardware so that sensitive data does not leave the company.
- A model is further trained to classify industry-specific documents.
What companies should keep in mind
- Choose appropriately: Not every task needs the largest model. Smaller models are often faster and cheaper.
- Check data processing: Clarify where requests are processed and which contract terms apply.
- Observe licenses: Freely available models also have license terms that may, for example, restrict commercial use.
- Limit dependencies: Models are developed further and discontinued. Applications should allow a model change without major rework.
- Know the regulation: The EU AI Act regulates general-purpose AI models separately. The obligations for their providers have applied since August 2, 2025. The Digital Omnibus Regulation (EU) 2026/1744 did not postpone this deadline.
How it differs
Foundation model describes a model's role as a broad basis. Large language model describes the type of model. Most well-known language models are both, but not every foundation model processes language, and not every language model is broad enough to count as a foundation model.