Large-scale AI models trained on broad data that can be adapted for a wide range of downstream tasks through fine-tuning or prompting. These models serve as versatile “foundations” upon which more specialised applications can be built.

Foundation models represent a paradigm shift in AI development by providing general-purpose capabilities that can be customised for specific use cases without training models from scratch. They encode general knowledge and patterns from diverse data sources, making them valuable starting points for specialised applications. While powerful, they also pose challenges related to interpretability, bias, and resource requirements. Major examples include BERT, GPT, T5, and CLIP.

An enterprise using a foundation model as the starting point for multiple AI applications, including document processing, customer support automation, and content creation, by fine-tuning the model on domain-specific data.