Definition
A foundation model is a model trained on broad data at sufficient scale to support multiple downstream uses. Applications can steer its behavior with prompts or adapt it through further training. Retrieval and tools extend what an application can do with the model, but belong to the surrounding system.
The term describes the model’s role as a reusable base. It does not prescribe one architecture, modality, or delivery method, although many current language foundation models use transformer architectures.
Simple example
A broadly pretrained language model can support summarization, information extraction, question answering, and code assistance. One team may use it with prompts and structured outputs for ticket classification. Another may fine-tune the same base model on reviewed support conversations to make its responses follow a particular style and task format.
Both applications start from capabilities learned during pretraining instead of training a model from scratch.
Why it matters
Foundation models shift much of model development from building a separate model for every task to choosing a shared base and adapting it to the task. That can reduce the data and compute needed for an application. Applications using the same base may also face the same underlying limitations and biases. Serving constraints and operational requirements depend on how the model is deployed.
Engineers still need to decide which behavior belongs in prompts, external retrieval, deterministic application code, or model adaptation.
One important nuance
Large and general-purpose are not precise thresholds. A model can be called a foundation model in one domain while remaining unsuitable for many tasks outside it. Broad pretraining also does not guarantee current or private knowledge.