Flower

Flower (the flwr package) is an open-source framework for federated learning: training one shared model across many machines without raw data ever leaving its owner. A bank, hospital, or phone fleet keeps its data local and only exchanges model updates, so the shared model learns from everyone's data while privacy and regulation are respected. It originated as a research project at the University of Oxford and is built to be customizable, extendable, and framework-agnostic, working with PyTorch, TensorFlow, Hugging Face Transformers, JAX, scikit-learn, and more.

The design favors readable, maintainable code and a gentle on-ramp: tutorials walk from federated-learning basics to custom strategies and clients. Note that Flower Labs' newer Collaborative Superintelligence and Flower Agent branding refers to a separate commercial product; the open-source flwr package documented here is the federated learning framework.

Honest limitations: federated learning is inherently harder than centralized training; debugging across distributed clients, handling stragglers, and securing aggregation are real engineering work. It also assumes your problem actually needs federation; if you can legally centralize the data, centralized training is simpler and faster.

pip install flwr and follow the quickstart to wire a server strategy to your client training code.