AI Potluck
Contribute
Back to Gap Map Model components / Federated learning

Flower

Flower Labs
open core / Overall score: 3.0

Federated learning framework from Flower Labs. The same app runs in its simulation engine or on its deployment runtime, where a SuperLink coordinates SuperNodes that each organization runs next to its own data over TLS with node authentication. It ships SecAgg and SecAgg+ secure aggregation and differential privacy, and works with PyTorch, TensorFlow, JAX, MLX and Hugging Face.

One product: `flwr-datasets`, a partitioning utility published from the same repository, is not declared, because its installs are Flower's own users counted a second time. Flower Enterprise features (OIDC account login, audit logging and the Helm charts) and the hosted SuperGrid service are Flower Labs' paid layer.

Openness

4 high confidence
4.0
license
Apache-2.0(OSI
source
public(the framework, simulation engine and deployment runtime)
core features withheld
yes — OIDC account authentication, audit logging and the Helm charts are documented as Flower Enterprise features, and the SuperLink reports license-missing and license-invalid errors

Flower's framework and deployment runtime are Apache-2.0, but Flower Labs reserves part of the production operating layer for a paid tier: its own documentation marks OIDC login, audit logging and the Helm charts as Flower Enterprise features. That is open core, not a separate product sold beside a complete one.

Adoption

2 medium confidence
2.0

PyPI downloads of `flwr`, which carries the simulation engine, the SuperLink and the SuperNode alike, so they count the product itself. Flower draws about two thirds of the monthly PyPI installs across this category.

Capability

4 high confidence
4.0

Flower runs real cross-organization federations, with authenticated SuperNodes, encrypted connections and secure aggregation in the open framework. It stops short of NVIDIA FLARE because its Android SDK does not work with current releases, its Slurm path runs through the hosted SuperGrid, and it publishes no multi-billion-parameter training figures.

Verified 2026-09-27