Foundation model
A foundation model is a large model trained broadly on general data, then adapted to many different tasks rather than built for one. The same model can be pointed at summarising, classifying, or conversing, which is what separates it from the task-specific models that came before.
Foundation model is the umbrella term. Large language models are the best-known kind, and the category also covers image, audio, and multimodal models. The word foundation refers to position in the stack rather than size: it is what everything else gets built on top of, and it is rarely the product a buyer actually purchases.
How do foundation models differ from earlier AI models?
Older machine learning models were trained for one task, on data labelled for that task. A model classifying support tickets learned from labelled tickets and could do nothing else, so a new task meant a new model and a new labelling effort.
A foundation model inverts that. It is trained once, expensively, on broad data, then adapted cheaply to many tasks through prompting, retrieval, or fine-tuning. The practical consequence is that capability now arrives before a use case is defined, which is why the same underlying model sits behind products that look nothing alike.
The cost profile shifted with it. Training is concentrated in a handful of organisations, and everyone else builds on top.
What do foundation models mean for a regulated buyer?
The thing you buy is rarely the model. It is an application built on a model somebody else trained. Questions that follow:
Which foundation models sit in the chain, and under whose agreement
Where each one processes data, and whether that satisfies your residency requirements
What happens when the vendor changes model, and whether you are told
Whether your evaluation results survive that change
Outsourcing does not reduce regulatory responsibility, and PRA SS2/21 is explicit on the point, so a model sitting in someone else's stack is still inside your control narrative.
Frequently asked questions
Is a foundation model the same as a large language model?
A large language model is one type of foundation model, the type that works with text. The broader category also includes image, audio, and multimodal models. In customer operations the two get used almost interchangeably, because the relevant foundation models are language models.
Can we train our own foundation model?
Realistically no, and almost nobody outside a handful of labs does. Training one costs enormously and needs data at a scale few firms hold. What is achievable is adapting an existing model with your own content and examples, which delivers most of the benefit at a fraction of the cost.
Does the choice of foundation model affect compliance?
It affects the questions you have to answer rather than the obligations themselves. Where the model runs, who processes the data, and what contractual terms apply all feed a control narrative. The duties on the firm stay the same whichever model sits underneath. Gradient Labs' seven questions to ask an AI agent vendor for financial services cover what to establish about the product sitting on top of the model, including whether compliance is built in or left for you to configure.