This paper provides a comprehensive survey of heterogeneous contrastive learning for foundation models, addressing the challenges posed by the exponential growth of big data. The authors categorize contrastive foundation models into two branches: those for view heterogeneity and those for task heterogeneity. They review the basic concepts of contrastive learning, including data augmentation, contrastive pair construction, and loss function formulation. The paper then delves into traditional multi-view contrastive learning methods and their application in training multi-view foundation models. It also discusses contrastive learning methods for task heterogeneity, including pre-training tasks and downstream tasks, and how different tasks are combined with contrastive learning loss for various purposes. Finally, the paper outlines several open challenges and future research directions in heterogeneous contrastive learning, such as representation redundancy and uniqueness, efficiency of CL models, better multi-view benchmark datasets, trustworthy CL, and understanding the mechanisms between CL strategies and downstream tasks.This paper provides a comprehensive survey of heterogeneous contrastive learning for foundation models, addressing the challenges posed by the exponential growth of big data. The authors categorize contrastive foundation models into two branches: those for view heterogeneity and those for task heterogeneity. They review the basic concepts of contrastive learning, including data augmentation, contrastive pair construction, and loss function formulation. The paper then delves into traditional multi-view contrastive learning methods and their application in training multi-view foundation models. It also discusses contrastive learning methods for task heterogeneity, including pre-training tasks and downstream tasks, and how different tasks are combined with contrastive learning loss for various purposes. Finally, the paper outlines several open challenges and future research directions in heterogeneous contrastive learning, such as representation redundancy and uniqueness, efficiency of CL models, better multi-view benchmark datasets, trustworthy CL, and understanding the mechanisms between CL strategies and downstream tasks.