AI glossary Models & Architecture
What is a foundation model?
← All glossary termsA large model trained on broad data that can be adapted, by prompting or fine-tuning, to many different tasks.
Explained
A foundation model is a large machine learning model trained on very broad data, usually with self-supervised learning, so that it picks up general patterns of language, images or code. Instead of training a new model for each task, teams adapt the same foundation model through prompting, retrieval or fine-tuning. The term was introduced by researchers at Stanford in 2021 to describe this shift. Large language models such as the GPT, Claude and Llama families are the best-known examples, and there are also foundation models for images, speech, code and other data types.
How a foundation model is built and used
- Pre-training: the model learns from a very large, broad dataset, for example by predicting the next word in text.
- Optional alignment: further training so it follows instructions and behaves safely.
- Adaptation: the same model is pointed at a task with a prompt, given company data through retrieval, or fine-tuned on examples.
- Deployment: it is used through an API or hosted on your own infrastructure, depending on the model and its licence.
Foundation model vs LLM
Every large language model is a foundation model, but not every foundation model is a language model. The term covers any large pre-trained model meant to be reused across tasks, including models for images, audio, video and multimodal input.
Because many applications share the same few models, their strengths and their flaws, such as biases or a tendency to state wrong facts confidently, are passed on to everything built on them.
What to weigh when choosing one
- Quality on your own tasks, measured with your own evaluation set rather than public benchmarks alone.
- Cost and latency at the volumes you expect.
- Hosting and data: where it runs, what happens to your inputs, and whether open-weight models are an option.
- Licence and regulatory obligations, including those for general-purpose AI models under the EU AI Act.
- Room to switch: designing the application so the model can be replaced as better or cheaper ones appear.
Frequently asked questions
What is an example of a foundation model?
The GPT, Claude, Gemini and Llama families of language models are foundation models. So are large image and speech models that are pre-trained once and then reused for many tasks.
Is ChatGPT a foundation model?
ChatGPT is a product built on top of foundation models from OpenAI. The underlying GPT models are the foundation models; the chat application adds the interface, tools and safety layers.
What is the difference between a foundation model and a generative AI model?
The terms overlap. 'Foundation model' describes how a model is built and reused (broad pre-training, many tasks); 'generative' describes what it does (produces new content). Most current foundation models are generative, but some are used mainly for classification or embeddings.
Do companies need to train their own foundation model?
Very rarely. Training one takes large amounts of data, compute and expertise. Most companies get better value by adapting an existing model with prompting, retrieval over their own data and, where it pays off, fine-tuning.
Why are they called foundation models?
Because they act as a common base on which many applications are built. The Stanford report that coined the term also stressed that this makes them a shared point of failure: weaknesses in the base carry over to everything built on it.
Decision guides that use this term
Where it shows up at Thinkia
Further reading
Related terms
Sources
- Bommasani et al. · On the opportunities and risks of foundation models (arXiv)
- Brown et al. · Language models are few-shot learners (arXiv)
- AWS · What are foundation models?
Updated: 2026-10-21