Examples
Worked examples
- Is an instance
A 70B-parameter LLM adapted by prompting to 30+ downstream tasks without further training.
- Is an instance
A vision-language foundation model fine-tuned to radiology, agriculture, and remote sensing applications.
Counter-examples
Looks similar, but isn't
- Not an instance
A task-specific small classifier (e.g., a logistic regression for spam).
- Not an instance
A handcrafted rule-based system.
Editorial commentary
A foundation model is a large machine-learning model trained on broad data at scale and designed to be adaptable to a wide range of downstream tasks through fine-tuning, prompting, or retrieval augmentation, rather than being trained for one narrow task from the outset. The term was coined and popularised by Stanford’s Center for Research on Foundation Models (CRFM) in a 2021 report, “On the Opportunities and Risks of Foundation Models” (Bommasani et al.), specifically to name a class of model that pre-dated the term — BERT and GPT-2/3 were already in this category — but had lacked a settled shared label.
How this differs from a frontier model — a genuinely contested distinction
“Foundation model” and “frontier model” are frequently used as if interchangeable, but they describe different axes, and the field has not settled a single operational boundary between them. Foundation model is a training-paradigm term: it describes how a model was built (broad pre-training, general-purpose adaptability), and that description applies permanently — a foundation model released years ago is still a foundation model today, however far behind current capability it now sits. Frontier model is a relative, capability-based or compute-threshold-based term describing where a model sits at a given point in time relative to the most capable systems then available — a model can lose its “frontier” status the moment a more capable system is released, without changing at all itself. Every frontier model is a foundation model, but most foundation models are not, and have never been, frontier models.
References
- Bommasani, R. et al. (2021). “On the Opportunities and Risks of Foundation Models.” Stanford CRFM.
- See also: Frontier model, Large language model (LLM), Open weights model.
Also known as
FM
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Foundation model"
vocab-term-identifier="https://casrai.org/dictionary/term/foundation-model" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/foundation-model",
"name": "Foundation model",
"identifier": "https://casrai.org/dictionary/term/foundation-model",
"description": "A large machine-learning model trained on broad data at scale and adaptable to a wide range of downstream tasks through fine-tuning, prompting, or retrieval augmentation.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/foundation-model",
"alternateName": [
"FM"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-09-05T14:24:40",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
