A foundation model is a large AI model trained on broad, general data that can be adapted — via fine-tuning, prompting, or retrieval augmentation — for a wide range of downstream tasks. GPT-4, Claude, Llama, and Gemini are all foundation models.
The "foundation" metaphor captures the idea that these models are starting points other applications build on, rather than training from scratch for every use case. Foundation model providers compete on benchmark scores, context length, safety, and price, while the application layer above them commoditizes rapidly as capabilities equalize.