Foundational Model

Foundational models learn general patterns and representations from large and diverse datasets during pre-training. Instead of being developed for only one specific task, they can support many applications through prompting, fine-tuning, or other adaptation techniques. Large language models and some multimodal models are common examples of foundational models.

Foundational models provide reusable AI capabilities that allow developers to build applications without training models entirely from scratch. Their broad capabilities can accelerate AI development, but their scale and widespread use also introduce considerations around security, bias, transparency, data privacy, and responsible deployment.