Flower is a friendly, open-source framework for federated AI — training and evaluating machine-learning models across many devices and data silos without moving the raw data. It lets you federate anything from small classical models to billion-parameter LLMs. The project has about 7,200 stars on GitHub and is actively developed in 2026.
In federated learning, the model travels to the data: phones, hospitals, banks or edge devices train locally and only share model updates, which a server aggregates into a global model. Flower makes this practical with a framework-agnostic core — it works with PyTorch, TensorFlow, JAX, scikit-learn, Hugging Face and more — and a runtime that scales from simulations on one machine to real deployments across Android, iOS and Raspberry Pi fleets. You define the client and strategy, and Flower handles communication, aggregation algorithms (FedAvg and many variants), privacy techniques such as differential privacy and secure aggregation, and evaluation. The project, stewarded by Flower Labs, also explores federated fine-tuning of large language models and federated analytics. Extensive documentation, tutorials and an active research community make it the default starting point for federated learning experiments and products alike.
Flower is free and open source under the Apache-2.0 license. You provide your own devices, servers and infrastructure.
Flower is for ML researchers and engineering teams in privacy-sensitive domains — healthcare, finance, mobile — who need to train models on distributed data they cannot centralize.