Apheris releases 3.9

The 3.9 release focuses on making federated computing easier to maintain and scale. It introduces a way to store and reuse model outputs, automates Gateway (Apheris' federated computing product) upgrades, and simplifies model development dependencies, all aimed at reducing repetitive setup work and improving experiment reproducibility.

Reusing model results without repeating computations

Training in federated setups often produces valuable results, e.g., model weights, checkpoints, metrics etc., that are hard to reuse or share later. The new Data Artifacts Store solves this by keeping those outputs organized in one place.

You can now:

• Save model weights and checkpoints automatically after each run

• Select previous results for downstream tasks without retraining

• Preserve experiment outputs to reproduce results later

• Control access: only the person who initiated the job can view or reuse its artifacts

This makes it easier to pick up work where you left off, continue from a specific training round, or validate results consistently across organizations.

Keeping Gateways up to date automatically

Previously, updating Apheris Gateways across organizations required manual work. Each Gateway had to be upgraded separately. With 3.9, Gateways running in on-premises or EKS environments can update themselves automatically. A built-in Flux controller now checks for new versions in a Helm repository and performs rolling upgrades when available. This reduces operational overhead, keeps all collaborators aligned on the same version, and prevents issues caused by version drift. The feature is optional and can be enabled via a feature flag when you’re ready to automate upgrades.

Linear Regression added to the regression model suite

The new Linear Regression model expands the available regression methods beyond Cox and Logistic Regression. It’s built to follow the same federated training interface, works seamlessly with the FederatedDataFrame, and supports use cases where a simple, interpretable regression is enough. For example, analysts can now train linear models on distributed datasets to understand variable relationships without deploying heavier models.

In a federated data network we did together with Roche (presented at ECTRIMS 2025), we saw a lot of success thanks to our statistics package handling of federated data frames. We found this functionality so useful that we decided to split it out to a more general-purpose package, referred to as the apheris_preprocessing package as of 3.9.

FederatedDataFrame becomes its own package

A FederatedDataFrame (FDF) is essentially a remote-preprocessing abstraction that behaves much like a pandas.DataFrame but is designed to work in a federated environment. It lets you define preprocessing steps (filtering, type casting, column operations) against datasets that sit behind separate Gateways (i.e., at different sites) without pulling the raw data into one place.

That preprocessing plan then runs at each collaborator's local data environment, so you maintain data location and privacy but still operate consistently across Gateways. With the 3.9 release of the Apheris Gateway, the FDF functionality has been pulled out into its own pip-installable package: apheris-preprocessing.

Key changes include:

• Better modularity: The preprocessing logic is now modularised: instead of being embedded inside the statistics or regression-models packages, FDF lives in its own package.

• Reduced Overhead: This means model packages (e.g., regression-models, model-xgboost) that rely on FDF no longer need to implicitly depend on the full statistics package.

• More flexible pipelines: For model and pipeline developers, this offers cleaner dependency graphs, lighter installations, and more flexibility to build custom federated preprocessing pipelines without pulling in unrelated statistical dependencies

Conclusion

The 3.9 release is all about making federated work easier to sustain and scale in real production environments. It gives teams more control over their model lifecycle, less maintenance overhead for infrastructure, and cleaner interfaces for developing and deploying models. With centralized artifact management, automated Gateway updates, and a modular preprocessing layer, this version lays the groundwork for more reproducible, maintainable, and flexible federated AI pipelines, whether used for clinical research, multi-institutional analytics, or privacy-preserving model training.

For detailed API documentation and migration notes, visit the Apheris documentation