A foundation model means a big, general AI model trained on a huge amount of varied data, which can then be adapted to many different jobs. Instead of building a new AI for every task, companies start from one foundation model and shape it for the job.
Think of the foundations of a house. They take the most work and money, and you cannot live in them as they are. But once they are poured, you can build a cottage, a shop or a block of flats on top. A foundation model is that groundwork: chatbots, writing tools and search features are the buildings put on it.
In a sentence: “We didn’t build our own AI: the help bot runs on a foundation model that we fine-tuned on our support articles.”
What the term means
The plain point: it is defined by two things: trained broadly, and adaptable to many tasks.
The name comes from a 2021 report by researchers at Stanford University. They described models “trained on broad data at scale” that “are adaptable to a wide range of downstream tasks”, and chose the word to show they are central but unfinished.1
Google’s glossary for developers says the same in two parts: a foundation model can respond well to a wide range of requests, and it can serve as a base for further training.2 That extra training is called fine-tuning.
A simple example
The plain point: one model, many products.
Amazon’s cloud business, AWS, explains that teams no longer build an AI from scratch: they use a foundation model “as a starting point”. It names BERT, a language model released in 2018, as one of the first.3 Companies use the phrase for their own models too. Apple calls the models behind Apple Intelligence its “foundation models”, including a small one that runs on the phone itself.4
Many AI tools share a handful of foundation models underneath. A strength or a flaw in that base can show up in every tool built on it.
Why it matters to you
The plain point: shared foundations mean shared problems.
The Stanford report warned that “the defects of the foundation model are inherited by all the adapted models downstream.”5 AWS lists answers that can be “unreliable” and sometimes “incorrect” among the challenges.3 So a polished app is no guarantee: check what matters, as hallucination meaning explains.
Where the word comes from
The term was introduced in August 2021 by a Stanford University research group. Its report, “On the Opportunities and Risks of Foundation Models”, came from the Center for Research on Foundation Models, which describes itself as “an interdisciplinary initiative” at Stanford’s Institute for Human-Centered Artificial Intelligence (HAI).6 The report explains the choice of word: “We call these models foundation models to underscore their critically central yet incomplete character.”7 Central, because many tools are built on them; incomplete, because they still have to be adapted for each job.
Often confused with
The plain point: three labels overlap, but each answers a different question.
| Word | What it describes | How it differs |
|---|---|---|
| Foundation model | Describes: how a model is used, as a general base for many tasks. | Covers: text, image and mixed models alike. The Stanford list included an image maker, DALL-E. |
| LLM | Describes: what a model works with: language. | Differs: most big LLMs are foundation models, but not every foundation model is about text. |
| Frontier model | Describes: a position: among the most capable right now. | Differs: California’s law defines a frontier model as a very large foundation model, so every frontier model is a foundation model but not the other way round. |
The California definition counts the computing power used in training.8
Questions people ask
Is ChatGPT a foundation model?
ChatGPT is an app. The GPT models inside it are foundation models; IBM calls them exactly that. See GPT meaning.9
What is the difference between an LLM and a foundation model?
“LLM” says what the model handles, language. “Foundation model” says how it is used, as a base for many jobs. A big chatbot model is usually both.
Learn more on this site
- How one is made: how an AI is actually built.
- Shaping it for a job: fine-tuning meaning.
- Who can afford to build them: who builds AI.
- The words: the AI glossary and AI model meaning.
- Related: What Is Pretraining in AI? Pretraining Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is a foundation model?
- A small model built for one single task
- A big, general model trained on varied data that can be adapted to many jobs
- The model that runs a building’s heating
Show the answer
Answer: A big, general model trained on varied data that can be adapted to many jobs. A foundation model is trained broadly and can be adapted to many tasks, so companies shape one for each job instead of starting again.
2. Why did the Stanford report choose the word ‘foundation’?
- Because the first ones were built by a charity foundation
- Because they only run on desktop computers
- To show they are central but still unfinished
Show the answer
Answer: To show they are central but still unfinished. The report says the word shows they are central, because many tools are built on them, and incomplete, because each job still needs adapting.
3. Why do shared foundations mean shared problems?
- A flaw in the base model can show up in every tool built on it
- Every app built on it gets slower over time
- All the tools must use the same name
Show the answer
Answer: A flaw in the base model can show up in every tool built on it. Many tools share a handful of foundation models, so a defect in the base is inherited by the tools built on top.
Related terms: Distillation · Emergent abilities · Fine-tuning · Frontier model · Inference · Latent space
- 1Stanford Center for Research on Foundation Models (Bommasani et al.), On the Opportunities and Risks of Foundation Models, arXiv:2108.07258, submitted 16 Aug 2021, read at source 5 Oct 2026“AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character.” ↑
- 2Google for Developers, Machine Learning Glossary, last updated 10 Apr 2026, read at source 5 Oct 2026“A very large pre-trained model trained on an enormous and diverse training set. A foundation model can do both of the following: Respond well to a wide range of requests. Serve as a base model for additional fine-tuning or other customization.” ↑
- 3AWS, What are Foundation Models?, read at source 5 Oct 2026“Rather than develop artificial intelligence (AI) from scratch, data scientists use a foundation model as a starting point to develop ML models that power new applications more quickly and cost-effectively.” … “Released in 2018, Bidirectional Encoder Representations from Transformers (BERT) was one of the first foundation models.” ↑a“Answers to questions on certain subject matter may be unreliable and sometimes inappropriate, toxic, or incorrect.” ↑b
- 4Apple Machine Learning Research, Updates to Apple’s On-Device and Server Foundation Language Models, 9 Jun 2025, read at source 5 Oct 2026“Our latest foundation models are optimized to run efficiently on Apple silicon, and include a compact, approximately 3-billion-parameter model, alongside a mixture-of-experts server-based model”. ↑
- 5Stanford Center for Research on Foundation Models (Bommasani et al.), On the Opportunities and Risks of Foundation Models, arXiv:2108.07258, read at source 5 Oct 2026“Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream.” ↑
- 6Stanford Center for Research on Foundation Models, Center for Research on Foundation Models (home page), read at source 12 Oct 2026“The Center for Research on Foundation Models (CRFM) is an interdisciplinary initiative” at Stanford HAI. ↑
- 7Stanford Center for Research on Foundation Models, On the Opportunities and Risks of Foundation Models (report page), 2021, read at source 12 Oct 2026“We call these models foundation models to underscore their critically central yet incomplete character.” ↑
- 8California Legislature, SB 53, Transparency in Frontier Artificial Intelligence Act, approved by Governor 29 Sep 2025, read at source 1 Oct 2026 (as cited on frontier model meaning): “‘Frontier model’ means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations.”↑
- 9IBM, What is GPT (generative pretrained transformer)?, read at source 5 Oct 2026“Developed by OpenAI, these foundation models power ChatGPT and other generative AI applications”. ↑
Stanford Center for Research on Foundation Models (Bommasani et al.), On the Opportunities and Risks of Foundation Models, arXiv:2108.07258, 16 Aug 2021, read at source 5 Oct 2026.
California Legislature, SB 53, read at source 1 Oct 2026 as cited on frontier model meaning. Applies in California, USA.
The comparison on this page (the foundations of a house) is this site’s own way of explaining, not a claim from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK