A foundation model is one big model that serves as the starting point for many products. Instead of training a separate model for translation, another for summaries and another for code, a lab trains one model on a huge, varied pile of data and then adapts it to each job. The term comes from a 2021 Stanford report, which described models trained on broad data at scale that can be adapted to a wide range of downstream tasks.
How it works
Training happens in two broad stages. First comes pretraining, where the model learns general patterns from enormous amounts of text, images, audio or code. The result is a base model that knows a lot but is not yet good at following instructions.
Then the model is adapted. That can mean fine-tuning on narrower data, post-training so it follows instructions and stays helpful, or simply writing a good prompt. The same foundation can end up as a chatbot, a coding assistant or a document reader.
Why it matters to you
Most AI tools you use sit on top of a foundation model. When you pick between assistants, you are mostly picking between foundations and the tuning on top of them.
Think of it like a general contractor who has learned carpentry, plumbing and wiring. You can hire that person for a kitchen or a bathroom. A specialist might do one room better, but the generalist can walk into almost any job.
That breadth cuts both ways. A flaw in the foundation, such as a bias or a gap in knowledge, is inherited by everything built on it. And a general model can sometimes match a purpose-built one without special training, which is why teams that maintain narrow models keep an eye on how quickly the general ones improve.
Many foundation models are large language models, but not all. Some handle images, audio or several kinds of input at once, which makes them multimodal models. The most capable ones are often called frontier models. A foundation model is a category defined by how it is used, while "frontier" describes how capable it is at a given moment.
You will also see the term in regulation. Laws and policy documents use "foundation model" or similar phrases to decide which models face extra rules, and the definitions differ from place to place. Some focus on training scale, others on how general the model's outputs are.
If a vendor says its product is "built on a foundation model," ask which one, how it was adapted, and what happens when the underlying model changes. That answer tells you where the quality, cost and risk really come from.