Definition
A foundation model is trained on broad data, generally through self-supervision at scale, and can be adapted to a wide range of downstream tasks. Adaptation may happen through prompting, retrieval, fine-tuning, adapters, tool use, or a task-specific model built on its representations.
The model is a foundation because other models and systems are built on it. That role creates leverage and shared dependence. An improvement in the base model can reach many applications, while a vulnerability, bias, licensing restriction, or undocumented training choice can propagate through the same chain.
Who introduced the term
Rishi Bommasani and coauthors introduced and defined "foundation model" in Stanford's 2021 report On the Opportunities and Risks of Foundation Models. They chose the name to describe the model's foundational role for downstream systems, not to claim that the model is the foundation of artificial intelligence itself.
The report applied the category beyond language to areas such as vision, robotics, and other data domains. The term therefore does not mean LLM, chatbot, or generative model, even though large language models are its most visible examples.
Distinguish it from nearby terms
- A large language model processes language tokens. It is a foundation model only when broad training and downstream reuse give it that role.
- A generative model produces new samples or content. A foundation model may be generative, discriminative, or support both uses.
- A fine-tuned model may descend from a foundation model while becoming specialized enough that it no longer serves as a broad base.
- AGI is a contested capability claim. Foundation-model status does not establish general intelligence.
Operational significance
Inventory the model's provider, version, license, training disclosures, modalities, adaptation method, and every downstream system that depends on it. A single upgrade can alter many products at once, so evaluation and rollback plans should follow the dependency graph rather than treat each application as isolated.
Check your understanding
A company fine-tunes a broad language model only for invoice coding. The base model is a foundation model; the resulting invoice classifier is a specialized descendant unless other tasks are intentionally built on it.