FedML™ (https://fedml.ai) provides a large open source community and a versatile enterprise platform for open, collaborative, and privacy-preserving machine learning (ML) over decentralized data and decentralized users. The key principle of FedML is to avoid any data pooling from edge devices (or data silos), and instead train local models directly where the data resides. These locally trained models would then be aggregated via FedML server to form a strong global model, which is then pushed back to the edge users. Compared with the conventional over-the-cloud ML pipeline, FedML avoids the critical challenges of data collection/pooling, slow adaptation/personalization of ML models to new conditions, and data privacy and security. It also eliminates the substantial cloud costs for ML development by leveraging the already available compute power at the edge.