Foundation Model

Last Updated: September 10, 2026 | By Mihail Sebastian | AI Dictionary

A large model pretrained on broad data and adapted to many downstream tasks rather than built for one, spanning language, vision, and multimodal AI.

What is a Foundation Model?

A foundation model is a large model pretrained on broad data at scale and then adapted to many downstream tasks, rather than built for a single one. Stanford researchers coined the term in 2021 to name a shift already underway: instead of training one model per task, organizations adapt one pretrained model to many.

The term is deliberately wider than “language model”. It covers text, image, audio, and multimodal systems, anything that serves as a base other applications build on.

How Foundation Models Work

A foundation model starts as a pretrained model: it learns general structure from massive unlabeled data, predicting missing or upcoming pieces of its input. No task labels are needed at this stage, which is what makes training at such scale possible.

Adaptation comes second, and it is the point of the exercise. The same base is specialized through fine-tuning on task data, through transfer learning, or simply through prompting, with no retraining at all.

This one-base-many-uses structure concentrates both capability and risk. A flaw in the base, a bias or a security weakness, is inherited by every application built on top of it.

Foundation Models vs Large Language Models

A foundation model is the broader class; a large language model is what you get when the modality is language. Every LLM is a foundation model, but a vision model adapted for medical imaging, satellite analysis, and quality inspection is a foundation model and not an LLM.

CriterionFoundation ModelLarge Language Model
What it namesThe broad class: any large pretrained model adapted to many tasksA large pretrained model for language
ModalityText, images, audio, video, or combinationsText, including code
RelationshipThe class itselfAn instance of the class
ExamplesLLMs plus vision and multimodal modelsGPT, Claude, Gemini, Llama

Applications of Foundation Models

Example of a Foundation Model

A provider releases one multimodal foundation model. A hospital fine-tunes it to draft radiology reports, a bank prompts it to pull figures out of scanned contracts, and a startup builds a customer-service agent on it.

Three products, three risk profiles, one base. Under the EU AI Act, that base is a general-purpose AI model: its provider must maintain technical documentation, give downstream developers the information they need to comply, and publish a summary of the training content, with extra evaluation and incident-reporting duties for models classified as posing systemic risk.

Related AI terms: Large Language Model · Pretrained Model · Transfer Learning · Multimodal AI · EU AI Act

Did you like the Foundation Model gist?

Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.

Written by

Mihail Sebastian

Mihail Sebastian

Editor, AI Guv

Mihail works in AI and writes about artificial intelligence topics for people who need to understand it without building it. He comes from more than 20 years of product design in startups.

Read the Governor's Letter

Stay ahead with Governor's Letter, the newsletter delivering expert insights, AI updates, and curated knowledge directly to your inbox.

By subscribing to the Governor's Letter, you consent to receive emails from AI Guv.
We respect your privacy - read our Privacy Policy to learn how we protect your information.

Browse All AI Terms A–Z

Every term in the dictionary, in alphabetical order. Jump to a letter or scroll the full list.

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z