Foundation model

What it means

A foundation model is a large AI model that a company trains once, on a very wide range of data, so that many other products can be built on top of it. The researchers who named it described these models as "trained on broad data at scale" and "adaptable to a wide range of downstream tasks."¹

Think of a building site. A builder pours a foundation first. On its own a foundation is not a home, but the same kind of base can hold up a house, a shop, or a school. In the same way, one foundation model can sit underneath a customer-service assistant, a coding tool, and an [AI agent](/definitions/ai-agent) that books travel. The researchers chose the word for exactly this reason: "a foundation is built first and it alone is fundamentally unfinished, requiring (possibly substantial) subsequent building to be useful."²

Where the word comes from

The term was coined at Stanford University. In August 2021, Rishi Bommasani, Percy Liang, and more than a hundred colleagues published a long report called On the Opportunities and Risks of Foundation Models. It said: "We call these models foundation models to underscore their critically central yet incomplete character."¹

In a follow-up post that October, Bommasani and Liang explained why they picked the name. They had turned down "(large) language model" because it "was too narrow given our focus was not only language." They also considered and set aside alternatives including "base model, broadbase model, inframodel, platform model, task-agnostic model, polymodel, pluripotent model, generalist model, universal model, multi-purpose model."²

Since then, the word has moved from research papers into law and marketing. In 2023 the U.S. government defined a narrower "dual-use foundation model" in an executive order. That order was revoked on January 20, 2025.³⁴ The European Union's AI Act took a different route and wrote its rules around a "general-purpose AI model" instead.⁵

How others define it

• Stanford researchers who coined it (2021): "models trained on broad data (generally using self-supervision at scale) that can be adapted to a wide range of downstream tasks."² "Self-supervision" is a training method; for a reader, what matters is that the model learns from enormous amounts of raw data.
• The EU AI Act (2024), under the name "general-purpose AI model": "an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications."⁵
• U.S. Commerce Department (NTIA), quoting the now-revoked executive order: "A dual-use foundation model means an AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts." The definition goes on to require high performance at tasks that "pose a serious risk to security."³

Where they agree and differ

All three agree on the core idea: one model, trained broadly, then reused for many purposes. They differ on scope. The Stanford definition is the widest. The EU adds that the model must show "significant generality" and be able to perform many tasks "competently."⁵ The U.S. definition added a size threshold and a link to serious risks. The Commerce Department noted that the wider AI community uses the term "notably without the 'tens of billions of parameters' requirement" and that "all foundation models do not necessarily display 'dual-use' capabilities."³

The strongest objection is to the word itself. Even its coiners admitted: "No name is perfect; even within the center, 'foundation model' is not everyone's favorite."² Europe's lawmakers chose "general-purpose AI model," which describes what the model does instead of using a building image.⁵ That objection holds up. Still, "foundation model" is the word executives will keep seeing in company announcements, so it is worth knowing.

In the news

On October 15, 2026, Microsoft AI said it had finished the first major training stage of MAI-Base-2, which it called "an MoE foundation model with 80 billion active and 2.8 trillion total parameters." (Parameters are the internal settings a model adjusts as it learns, and are a rough measure of size.) Microsoft said this "will enable us to license our model weights to trusted partners through Foundry."⁶ These are Microsoft's own claims and have not been independently checked. We are confident about what Microsoft said, not about how well the model performs.

The same day, Mistral released "our first family of time-series foundation models," which are built to forecast numbers over time rather than to write text.⁷ Together the two announcements show how far the word has spread: it now covers any broadly trained base that others build on, not just language models.

What it means for your organization

Most organizations will never build a foundation model. They will rent one, often without knowing it, inside the tools and agents they buy. The Stanford authors warned that "the defects of the foundation model are inherited by all the adapted models downstream."¹ If several of your vendors' agents run on the same foundation model, a flaw in it can show up in all of them at once. The model is the bottom layer. What your agents know about your organization, and the limits they work within, is something you supply on top of it (see [knowledge core](/definitions/knowledge-core)).

One question to ask your team: Which foundation models do our AI tools and agents actually run on, and what happens to us if one of them changes or fails?