Federated learning as a platform
The model travels to the data. Not the other way around.
Train your ML model across thousands of installations without moving a single row of data. No ingestion pipelines, no storage bills, no privacy risk. Shipped as one import line or a Docker container.
What you actually get
Money you don't spend. Time you don't lose.
No ingestion pipeline to build. No data lake to pay for. No release cycle to wait on. You upload a model, and you watch it get better across every installation from one dashboard.
Money saved
No terabytes to upload, no central storage bill, no ingestion pipeline to build and keep alive. The infrastructure you would have needed is simply not part of the design.
Time gained
Integration is one import line or one container, not a specialist project measured in months. An improved model reaches your customers without a release.
Risk removed
Training data never leaves the customer's machine, so the privacy conversation is a short one. Data you were never allowed to pool becomes trainable.
How it works
The data stays. The knowledge grows.
Federated learning sounds complicated. The principle is simple, explained in four steps.
01
Your data stays home
Your customers' data never leaves their devices or servers. No upload, no central data pool, no risk.
02
The model learns locally
During scheduled training windows, every installation trains your model on the data where it lives. At night, when idle, without disturbing operations.
03
Only the learning travels
Instead of data, only tiny model updates are sent to the cloud. Encrypted, anonymized, and orders of magnitude smaller than the data itself.
04
Everyone benefits
The server in the cloud merges the updates and distributes the improved model back to everyone. Each participant gains from the knowledge of all, without revealing anything.
The platform
Everything around your model. Already built.
You bring the model. The platform handles the rest: orchestration, versioning, rollout, and observability across your entire fleet.
Upload wizard
Upload your model, confirm the detected training functions, done. No rewriting of your code.
Training windows
You decide when training runs. For example at night from 2 to 4, with an audit log.
Promote & rollback
New model versions go live only when you decide. Rollback is one click.
Fleet overview
Which installation serves which version, who trained, who is unhealthy. Anonymized.
Ready-made artifacts
Python SDK or Docker container, preconfigured with credentials. Download and ship.
Local inference
The artifact doesn't just train, it serves the model locally. Updates roll out automatically.
Data sovereignty
Here, sovereignty isn't an option. It's the architecture.
01
Data never leaves the client
Not for training, not for predictions. Only encrypted model updates travel, orders of magnitude smaller than the data.
02
Scheduled and auditable
Training runs in windows you define. Every window, every round, every model version is logged.
03
Anonymous client IDs
Every installation gets a random identifier. The mapping to real locations exists only on your side, if you keep one at all.
04
Operated in the EU
Platform and model storage run in the Frankfurt region, encrypted at rest and in transit.
At no point do we hold any mapping between client IDs and real persons or locations. We could not de-anonymize your users even if we were ordered to.
Pricing
Fair, by the size of your fleet.
Start free. Pay only when your federation grows. Pricing scales with the number of clients and models you run, not with how much you train.
Free
1 model
up to 2 clients
- Full platform feature set
- Access to the community federation
- Operated in the EU
Node
3 models
up to 100 clients
Everything in Free, plus:
- Promote & rollback, training windows
- Fleet overview and round history
- Email support
Cluster
10 models
up to 1,000 clients
Everything in Node, plus:
- Audit logs and fleet analytics
- Alerts on failed training windows
- Priority support
Constellation
Unlimited models
Unlimited clients
Everything in Cluster, plus:
- SLA and dedicated support
- SSO and advanced roles
- On-premise options on request
* Indicative. Final tiers on request. Training is included under fair use on every plan.
The community federation
Coming soonPublish your model to the community: others use it for free, and in exchange their clients train it for you. Your model grows faster than you could ever collect data alone. Selected premium models can be offered for a fee.
You give your model
↓
Their clients train it for you
For your engineers
One import line. Or one container.
Your engineering team integrates the platform in minutes, not weeks. Two paths, same result.
# install
pip install flp-client
# connect and start
import flp
client = flp.Client(enrollment_key="flp_ek_…")
client.serve() # trains in windows, serves predictions locally
prediction = client.predict(sample)Both paths include training, local inference, and automatic model updates. For apps without Python, the container exposes a local REST API.
FAQ
Frequent questions
Everything the Flower ecosystem supports: PyTorch, TensorFlow, scikit-learn, XGBoost and more. Your model only needs to provide two functions, training and evaluation. Our upload wizard usually finds them automatically.
Honest answer: the model weights live on every client, that is how federated learning works. Treat the architecture like shipped code. What stays protected are your customers' data and the central training history.
If a client can train the model, it can also serve it. Training is the demanding part and runs in scheduled windows. The upload wizard checks your model's feasibility before rollout.
Not yet. Version 1 supports everything Python or Docker runs on: servers, edge devices, workstations. Mobile clients (iOS, Android) are on the roadmap.
There is a free tier with one model and a limited number of clients. Beyond that you pay monthly, tiered by the number of clients. See the pricing section for details.
The architecture is built for it: data stays with the data subject, there is no central data pool, client IDs are anonymous to us, and operations run in the EU. Responsibility as the data controller remains with you; we provide the technical foundation and the documentation for it.
Join us
Join the federation.
Tell us briefly what you have in mind. We'll set up your access and get back to you within two business days.
Stay in the loop
Product updates and insights on federated learning. Rare, but substantial.
Unsubscribe anytime. Never shared with third parties.