What is a foundation model?
A vendor pitches a custom-trained model at a price that assumes you need one. Most buyers in that meeting need a wrapper around GPT or Claude, not a model built from scratch.
Foundation model
base model
A foundation model is a large model trained once on broad, general data. It then gets adapted to many downstream tasks through prompting, retrieval or fine-tuning, instead of being trained from scratch for each one. GPT-4, Claude and Gemini are all foundation models: a base other systems get built on.
Stanford's Center for Research on Foundation Models coined the term in 2021. One model trained at large scale, then adapted to a wide range of downstream tasks rather than purpose-built for one. Adaptation happens three ways: write a better prompt, retrieve facts and pass them in, or, less often, fine-tune the model on your own examples.
Large language model (LLM) is a narrower term. Every LLM is a foundation model, because text prediction at scale is what made the category possible. Not every foundation model is only an LLM: some, like GPT-4o and Gemini, take images and audio too. "Foundation model" is the wider label; "LLM" names its best-known and most common species.
You are almost never training one
OpenAI, Anthropic and Google spend on compute and training data that no single product budget matches. That is the whole economic point of a foundation model: one company absorbs that cost once, and everyone else rents the result. Building your own from scratch is not a decision most teams ever face, because the market already solved it.
So the real question is not which foundation model to train. It is which one to build on, and what to wrap around it. Two products can call the same API and behave nothing alike, because the differentiation moved. It now lives in the system prompt, the retrieval pipeline, the guardrails, and the evals that catch a bad answer before a customer sees it.
That is the business we are in. We do not train foundation models. We build the layer around one that decides whether it holds up in production. Retrieval grounds it, checks catch it when it is wrong, and support keeps it working after the model provider ships a new version.
- Foundation modelTrained once, broadly, by a lab.
- AdaptPrompt, retrieve, or fine-tune.
- GroundYour data, your rules.
- GuardChecks before a user sees output.
- ProductWhat the buyer actually pays for.
Four of these five steps are ours to build. The first almost never is.
Related questions
01Is a foundation model the same as an LLM?
No. Every LLM is a foundation model, but a foundation model can also handle images, audio or video, which an LLM by definition does not. LLM names the text-prediction species; foundation model names the wider category it belongs to.
02What are some examples of foundation models?
GPT-4 and GPT-4o from OpenAI, Claude from Anthropic, and Gemini from Google are all foundation models. Meta's Llama family and Mistral's models are open-weight foundation models you can self-host instead of calling through an API.
03Do I need to train my own foundation model?
Almost never. Training one from scratch requires compute and data budgets that only a handful of labs carry. Most products adapt an existing foundation model through prompting, retrieval or fine-tuning instead, which is faster and far cheaper to maintain.
04How is a foundation model different from fine-tuning?
A foundation model is the starting point. Fine-tuning is one way to adapt it. It trains the model further on your own examples, so it behaves a certain way by default without learning new facts.
