What is a foundation model?

The plain idea, and the name

Traditionally, an AI model was trained for one specific task — a spam classifier, a translation model, an image tagger — and was not much use outside that task. A foundation model breaks that pattern: it is trained once, at great scale, on broad and general data, and the result is a capable, general-purpose starting point rather than a narrow tool.

The name reflects that role. It is not the finished application a user interacts with; it is the foundation underneath many different applications, each built by adapting the same underlying model to a different purpose. One foundation, many things built on it.

How it relates to LLMs and generative AI

A large language model is one kind of foundation model — one focused on text, trained broadly enough to write, summarise, answer and converse across an enormous range of subjects. Our explainer on large language models covers what makes them work; this article deliberately does not repeat that ground.

Foundation models are not limited to text, though. There are foundation models built around images, and multimodal foundation models that work across text, images and other data together. Generative AI, covered in our explainer on what generative AI is, is the broader category of AI that creates new content — and foundation models, including LLMs, are the technology that generative-AI applications are typically built on.

How you adapt one to your needs

The point of a foundation model is that you rarely train one from scratch — that is a huge undertaking reserved for the organisations building the model itself. Instead, you adapt an existing foundation model to your task, usually in one of three ways: prompting, where you simply instruct the model well (our explainer on prompt engineering covers this); retrieval-augmented generation, where you supply the model with relevant documents at the moment of answering (our explainer on RAG covers this); or fine-tuning, where you train the model further on your own examples to shift its behaviour (our explainer on fine-tuning covers this). All three adapt a foundation model without building a new one.

Why foundation models changed AI

The practical shift is significant: instead of every team training its own narrow model for its own task, one capable foundation model can be adapted to many tasks, which is faster, cheaper and more broadly capable than the old one-model-per-task approach. Most organisations never train a foundation model themselves — they consume one through an API or a managed cloud service, adapting it with prompting, retrieval or fine-tuning rather than building infrastructure to train models at that scale. Amazon Bedrock, covered in our explainer on Amazon Bedrock, is the AWS expression of this — a managed service offering access to foundation models without requiring you to run the underlying training or hosting infrastructure yourself.

The honest limits

A foundation model inherits whatever biases and gaps existed in its broad training data — being general-purpose does not make it neutral, a concern covered in our explainer on responsible AI. It can still hallucinate, generating fluent but incorrect content, and it has a knowledge cutoff — a point in time after which it simply was not trained on anything, which is exactly the gap that retrieval-augmented generation exists to close for time-sensitive or private information.

Being broad and reusable also does not mean a foundation model is equally good at everything; adapting it well for a specific task still takes real effort, whether that is careful prompting, a retrieval layer, or fine-tuning.

Foundation models in certification study

The concept of a foundation model is central to the AWS Certified AI Practitioner (AIF-C01), which expects candidates to understand what a foundation model is, how it differs from a task-specific model, and how it is adapted through prompting, RAG and fine-tuning — the same three routes covered above. For a broader architectural view of where these models fit in a cloud deployment, the AWS Certified Solutions Architect – Associate touches on the surrounding services; for building applications on top of one, the AWS Certified Developer – Associate is the natural next step.