BETA
A foundation model is a large model pre-trained on broad data that serves as a base for many downstream tasks through prompting or fine-tuning.
The term was coined by Stanford researchers in 2021 to capture a shift: instead of training a separate model per task, one massive pre-trained model is adapted — via prompting, fine-tuning, or distillation — to power thousands of applications. Foundation models now span text, code, images, audio, video, and scientific domains like protein structure.
Providers typically release a foundation model as a family: a base version for further training and instruct versions ready for chat and API use.
The economics are the point. Pre-training costs tens to hundreds of millions of dollars, but that cost is amortized across every downstream use, which is why renting foundation-model capability per token is dramatically cheaper than training anything comparable yourself.
Last revised 2026-07-05 · All glossary terms → · Live AI model pricing →