Oríon
Oríon is a free, open-source framework for asynchronous black-box function optimization, aimed at tuning machine-learning models and training.
What Oríon is
Oríon is a free, open-source framework for asynchronous black-box function optimization, aimed at tuning machine-learning models and training. It can work with command-line scripts with minimal workflow changes: users define parameter search spaces, configure runs through command-line arguments or configuration files, and report objectives through the Python client API or report_objective(). Included algorithms are Random Search, Hyperband, TPE, and ASHA, with additional algorithms available through plugins. Oríon supports asynchronous parallel optimization using documented Joblib and Dask worker backends. Trials are stored in EphemeralDB, PickledDB, or MongoDB, and experiment versioning records trials and parameters for reproducibility. PickledDB is generally unsuitable for parallel optimization above 50 workers; MongoDB is recommended beyond that scale. Visualizations include regret curves, parallel coordinates, local parameter importance, and partial dependencies. Oríon runs on Linux, macOS, and Windows, installs with pip install orion, supports self-hosting and API access, and uses the BSD license.
Good
- Supports asynchronous parallel optimization
- Includes four listed optimization algorithms
- Accepts command-line and configuration-file setup
- Provides experiment versioning and visualizations
- Offers API and self-hosted deployment
Watch out
- PickledDB is unsuitable above 50 parallel workers
- Larger parallel runs require MongoDB recommendation
- Not a machine-learning framework or pipeline
Who it suits
It suits machine-learning researchers and engineers tuning models or training within existing scripts. It is not itself a machine-learning framework, pipeline, or automated machine-learning product.
Verdict
Oríon adds configurable optimization, storage, comparison, and visualization to existing research workflows. Its database guidance matters when scaling parallel experiments.
The facts
Each one read from Oríon's own pages, with the date it was checked.
What it does
- Algorithms
- The installation guide lists Random Search, Hyperband, TPE, and ASHA as included algorithms, with more available through plugins.checked 4 Oct 2026
- Database backends
- The documented built-in backends are EphemeralDB, file-based PickledDB, and MongoDB; MongoDB is recommended for large-scale parallel optimization above 50 workers.checked 11 Oct 2026
- Experiment tracking
- Oríon stores trials in a configured database and offers experiment versioning to track trials and their parameters for reproducibility.checked 11 Oct 2026
- Installation
- The installation guide directs users to install Oríon from PyPI with pip install orion.checked 11 Oct 2026
- Integration
- A script can report its optimization objective through Oríon’s Python client API.checked 4 Oct 2026
- Intended users
- The documentation describes Oríon as useful for machine learning researchers and engineers, while saying it is not itself a machine learning framework, pipeline, or automated machine learning product.checked 11 Oct 2026
- License
- The project documentation says Oríon is licensed under the BSD license.checked 4 Oct 2026
- Optimization algorithms
- The documentation lists Random Search, Hyperband, TPE, and ASHA as included algorithms, and supports adding algorithms through plugins.checked 11 Oct 2026
- Parallel execution
- Oríon is natively asynchronous and the documentation says it can run in parallel from a laptop to a computing farm with thousands of processors.checked 11 Oct 2026
- Purpose
- Oríon is an asynchronous framework for black-box function optimization, including hyperparameter optimization for machine learning.checked 11 Oct 2026
- Scaling limit
- The documentation says PickledDB is generally not suited to parallel optimization with more than 50 workers, though it describes this as a rule of thumb.checked 11 Oct 2026
- Search spaces
- Oríon supports search spaces defined by priors, including almost all SciPy distributions, configured through the command line or a configuration file.checked 11 Oct 2026
- Storage
- The documented database backends are in-memory EphemeralDB, file-based PickledDB, and MongoDB.checked 4 Oct 2026
- Storage limit
- The documentation says PickledDB is generally not suited to parallel optimization with more than 50 workers; MongoDB is recommended above that scale.checked 4 Oct 2026
- Support
- The project directs users to report questions, issues, bugs, or feature suggestions through its GitHub issue tracker.checked 11 Oct 2026
- Visualization
- The documentation includes experiment visualizations such as regret curves, parallel coordinates, local parameter importance, and partial dependencies.checked 4 Oct 2026
- Workflow
- It can tune command-line scripts with minimal changes to a researcher’s existing workflow.checked 4 Oct 2026
- Workflow integration
- The documentation says users can integrate Oríon into an existing workflow by calling report_objective() at the end of their script, or use its Python API.checked 11 Oct 2026
Where it runs
- Platforms
- The installation guide says Oríon supports most Linux distributions and Mac OS X, and Windows 10 through Windows Subsystem for Linux.checked 11 Oct 2026
Alternatives to Oríon
- 1
Weights & Biases
Free planAndroidiOSWebAPI8.3From $60 a month - 2
Comet
Free planWebAPILinuxmacOS7.9From $19 a month - 3
DagsHub
Free planWebAPILinuxmacOS7.9From $99 a month - 4
Metaflow
Free planOpen sourceWebAPILinuxmacOS7.5Free plan - 5
ClearML
Free planWebAPILinuxmacOS7.4From $15 a month - 6
ZenML
Free planWebAPILinuxmacOS7.1From $999 a month