Foundation models are large-scale AI models trained on diverse, unlabeled data using self-supervised learning. They can perform a wide variety of downstream tasks, like text generation, summarization, image analysis, or code generation, making them highly reusable across applications.