Before foundation models, building an AI system for a specific task usually meant training a model from scratch on data specific to that task. A foundation model flips this: it's trained once on a very broad dataset, and that single trained model becomes the base for a huge range of downstream applications, either used directly or lightly adapted for a specific use case.

This is why the same underlying model can power a customer service chatbot, a coding assistant, and a document summarizer, the foundation model provides broad general capability, and the specific application is built on top through prompting, fine-tuning, or connecting it to particular tools and data.

Frequently Asked Questions

Why is it called a "foundation" model?

Because it serves as the base layer that many different, more specific applications are built on top of, the same way a building's foundation supports many different structures above it, rather than being built for one narrow purpose.

What's the difference between a foundation model and a large language model (LLM)?

LLMs are the most common type of foundation model, trained on text. Foundation model is the broader category, and can include models trained on images, audio, or multiple modalities together, not just text.

How does a company adapt a foundation model for its own specific use case?

Common approaches include prompting (giving it detailed instructions and context at query time), retrieval-augmented generation (feeding it relevant company-specific data), and fine-tuning (further training it on company-specific examples), each with different cost and complexity tradeoffs.

Do all AI agents rely on a foundation model underneath?

Nearly all modern AI agents use a foundation model as their reasoning core, with agent-specific logic (tool calling, memory, orchestration) built around it to turn that general reasoning capability into task completion.