A foundation model is trained on broad data at scale and can support a range of applications rather than one narrowly defined output. Language, image and multimodal models can serve as foundations for assistants, search, classification and content-generation systems.
For a builder, choosing a foundation model usually means choosing a starting point: prompt it as is, add retrieval for your own documents, or fine-tune it for a narrow job. Because many products depend on a small number of such models, plan for change by keeping prompts and LLM evaluation sets portable between providers.
The label does not guarantee that a model is reliable, unbiased or suitable for a particular use. Evaluate its capabilities, limitations, licensing, privacy implications and cost in the context of the application you are building. Related: Large Language Model, Multimodal AI.