Foundation Models
Foundation models are large models pretrained on broad data that can be adapted to many downstream tasks.
What Is a Foundation Model?
Before foundation models, building an AI system for a specific task usually meant training a model from scratch on data for exactly that task. A foundation model flips this: it's trained once, on a huge and broad dataset, to build general capability — and that single trained model becomes the starting point for many different downstream applications, adapted through prompting, RAG, or fine-tuning rather than training from zero each time.
Foundation Model
Trained OnceOne broadly-capable model, built one time.
Chat Assistant
One AdaptationAdapted through prompting, RAG, or fine-tuning.
Code Generation
Another AdaptationThe same underlying model, a different use.
Summarization
Another AdaptationNo training from scratch needed.
Classification
Another AdaptationYet another downstream task, same starting point.
Key Idea
"Foundation" describes the role a model plays — a broad, general-purpose starting point — not a specific architecture or size.
Why This Shift Mattered
- Reuse — the same underlying model can power a chat assistant, a summarizer, and a classifier, instead of needing a separately trained model for each.
- Lower barrier to entry — most teams building AI applications today adapt an existing foundation model rather than training one from scratch, which most organizations couldn't realistically do.
- Emergent general capability — training on broad data at scale gives foundation models abilities — like following instructions on tasks they weren't specifically trained for — that narrower, task-specific models don't develop.
Important
An LLM is a foundation model specialized for language — the terms overlap heavily in practice, but "foundation model" is the broader category, and includes models for images, audio, and other modalities too.
Common Mistakes
Treating "foundation model" and "LLM" as identical
"Foundation model" is the broader category — LLMs are foundation models specialized for text, but the term also covers models for images, audio, and other modalities.
Assuming every AI application needs to train its own foundation model
The entire point of a foundation model is that it can be adapted by many downstream applications without each one retraining from scratch.
Assuming a foundation model works equally well on any task out of the box
Fine-tuning, prompting, or RAG are often still needed to adapt a general-purpose foundation model to a specific task well.
Interview Question
What is a foundation model, and why did it change how AI applications get built?
A foundation model is trained once, on a broad dataset, to build general capability — and that single model then becomes the starting point for many different downstream applications, adapted through prompting, RAG, or fine-tuning instead of training from scratch each time. This changed how applications get built because most teams today adapt an existing foundation model rather than training their own, which most organizations couldn't realistically do at that scale. An LLM is a foundation model specialized for text, but the broader category also includes models for images, audio, and other modalities.
What an interviewer may ask next
- How is a foundation model different from a model trained for one specific task?
- Is an LLM the same thing as a foundation model?
- Why doesn't every team need to train their own foundation model?
Explain It in 30 Seconds
A foundation model is trained once, on broad data, to build general capability, and becomes the starting point many different applications adapt through prompting, RAG, or fine-tuning, rather than each training a model from scratch. This is why most AI applications today are built on top of an existing foundation model instead of training one. An LLM is a foundation model specialized for text; the broader category also covers other modalities like images and audio.