In plain words
A foundation model is trained on broad data at scale and can support a range of downstream applications. The term emphasizes its role as a reusable base rather than a system built for just one narrowly defined task.
A closer look
Adaptation can happen through prompting, fine-tuning, retrieval, or integration with tools. A language foundation model may power summarization, drafting, classification, and coding. Other foundation models work with images, audio, video, or multiple kinds of input.
Reusing a common base can make development easier, but it can also spread the same weaknesses across many applications. Training-data limitations, security issues, or systematic errors may appear in several products built on the same model. The final application still needs its own evaluation.
In practice
Two organizations use the same pretrained model: one builds a writing assistant, while the other develops a document classifier. Their interfaces and adaptation methods differ, but they share a foundation.
A useful distinction
“Foundation” is not a quality certification, and it does not imply general intelligence. A broadly capable base model can still fail at a specific task, require substantial adaptation, or be unsuitable for a particular use.