---
title: "What is a foundation model in AI?"
description: "A foundation model is a large AI model trained on broad data that can be adapted to many tasks. Where the term came from, and how it differs from an LLM."
type: "glossary-term"
canonical_url: "https://theaiengineering.com/glossary/foundation-model"
markdown_url: "https://theaiengineering.com/glossary/foundation-model.md"
topic: "AI engineering"
author: "The AI Engineering Editors"
publisher: "The AI Engineering"
updated: "2026-11-06"
keywords: ["Foundation model","foundation models"]
citation: "The AI Engineering Editors. \"What is a foundation model in AI?\" The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/foundation-model"
---

# What is a foundation model in AI?

Also called foundation models.

> **Definition of Foundation model:** A foundation model is a large AI model trained on broad data, usually with self-supervised learning at scale, that can be adapted through prompting or fine-tuning to a wide range of downstream tasks.

A foundation model is a model trained on broad data at scale that can be adapted to a wide range of downstream tasks. One foundation model can sit behind a chatbot, a classifier, a code assistant, and a summarizer, adapted to each through prompts, retrieved context, or [fine-tuning](https://theaiengineering.com/glossary/fine-tuning). That reuse is what separates AI engineering from earlier machine learning work, where teams trained a separate model for each task. For an AI engineer, the foundation model is a dependency: chosen, adapted, and evaluated, but rarely trained.

## Where the term foundation model comes from

Stanford's Center for Research on Foundation Models (CRFM) introduced the term in the 2021 report [On the Opportunities and Risks of Foundation Models](https://arxiv.org/abs/2108.07258) (Bommasani et al.). The report describes models trained on broad data, generally with self-supervision at scale, that can be adapted to many downstream tasks, and names BERT, DALL-E, and GPT-3 as examples.

In a follow-up post, [Reflections on Foundation Models](https://crfm.stanford.edu/2021/10/18/reflections.html), Rishi Bommasani and Percy Liang explained the name. They rejected "large language model" as too narrow, because the same pattern applies to images, code, proteins, and speech. They rejected "pretrained model" because it implied that everything important happened after pretraining, and "self-supervised model" because it named only the training objective. A foundation is built first and is unfinished on its own, which matches how these models are used.

The report also names the main risk of the pattern, homogenization. When many applications adapt the same model, a defect in that model is inherited by every application built on it.

## Foundation model vs large language model

A [large language model](https://theaiengineering.com/glossary/large-language-model) is one kind of foundation model. "Foundation model" describes a role: a general model trained once and adapted many times, whatever its data type. "Large language model" describes a model that reads and generates text tokens. Every general-purpose LLM is a foundation model, but image, speech, and protein models can be foundation models too.

| | Foundation model | Large language model |
|---|---|---|
| Defined by | Broad training data and adaptability | Generating text, one token at a time |
| Data types | Any: text, images, audio, code, proteins | Text and code, sometimes with image or audio input |
| Examples | GPT-3, BERT, DALL-E, CLIP | GPT-3 |

[CLIP](https://arxiv.org/abs/2103.00020) (Radford et al., 2021) shows a foundation model outside text generation. It was pretrained on 400 million image and text pairs collected from the internet, and the authors benchmarked its zero-shot transfer on more than 30 existing computer vision datasets, where it was often competitive with fully supervised baselines. [Multimodal models](https://theaiengineering.com/glossary/multimodal-model) build on encoders of this kind.

## How applications adapt a foundation model

Applications adapt a foundation model in three main ways, from cheapest to most involved: prompting, which changes the instructions; retrieval, which adds facts to the context at request time; and fine-tuning, which changes the weights. All three reuse the expensive [pretraining](https://theaiengineering.com/glossary/pretraining) run, which is why one model can serve many products.

Teams also choose how to reach the model: through a provider's API for proprietary models, or by downloading an [open-weights model](https://theaiengineering.com/glossary/open-weights-model) and serving it themselves. Homogenization applies here too. A provider's model update reaches every feature built on that model at once, so production teams pin model versions and rerun their [evaluation](https://theaiengineering.com/glossary/llm-evaluation) suite before switching.

[What is AI engineering](https://theaiengineering.com/articles/what-is-ai-engineering) explains the layers that teams build between a foundation model and a product.

## Related terms

- [Large language model (LLM)](https://theaiengineering.com/glossary/large-language-model): A large language model (LLM) is a neural network, usually a transformer with billions of parameters, trained on web-scale text to predict the next token and then adapted to follow instructions.
- [Pretraining](https://theaiengineering.com/glossary/pretraining): Pretraining is the first stage of training a language model, in which it learns to predict the next token across a large unlabeled corpus of text and code, producing a general-purpose base model.
- [Fine-tuning](https://theaiengineering.com/glossary/fine-tuning): Fine-tuning is the process of further training a pretrained model on a smaller, task-specific dataset so its weights change to improve performance on a particular task, domain, style, or output format.
- [Open-weights model](https://theaiengineering.com/glossary/open-weights-model): An open-weights model is an AI model whose trained parameters are published for download, so anyone can run, fine-tune, or quantize it on their own hardware under the terms of its license.
- [Multimodal model](https://theaiengineering.com/glossary/multimodal-model): A multimodal model is an AI model that takes in or produces more than one type of data, such as text, images, audio, or video, and combines them within a single model.
- [Transformer](https://theaiengineering.com/glossary/transformer): A transformer is a neural network architecture that processes sequences with self-attention, letting each token weigh information from other tokens in parallel; it is the basis of most modern large language models.

## Guides that cover foundation model

- [What is AI engineering? The discipline, the stack, and the work](https://theaiengineering.com/articles/what-is-ai-engineering): What is AI engineering? A plain definition, how it differs from ML and software engineering, the layers of the stack, and what AI engineers do day to day.
- [AI engineer roadmap: skills, projects, and resources by stage](https://theaiengineering.com/articles/ai-engineer-roadmap): A practical AI engineer roadmap in seven stages, each with the skills to learn, a project to ship, and resources worth the time, from Python to production.

Part of the [AI engineering](https://theaiengineering.com/topics/ai-engineering) topic hub.

## Cite this definition

The AI Engineering Editors. "What is a foundation model in AI?" The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/foundation-model

- Title: What is a foundation model in AI?
- Author: The AI Engineering Editors
- Publisher: The AI Engineering
- Updated: 2026-11-06
- URL: https://theaiengineering.com/glossary/foundation-model
- Markdown: https://theaiengineering.com/glossary/foundation-model.md
