Polyaxon is an open-source AI control plane for Kubernetes: a platform for running everything from GPU training jobs to agent sessions on your own infrastructure, with scheduling, tracking and governance in one place. Teams use it to train models, run hyperparameter sweeps and pipelines, and serve workloads without giving up control of their clusters and data. The project has around 3,700 stars on GitHub, is Apache-2.0 licensed, and is actively developed.
Polyaxon treats training jobs, services and sandboxes uniformly: every run is described by a spec, queued under quota controls, and recorded with full lineage of code, data and artifacts. It supports distributed training frameworks such as PyTorch DDP, MPI and Horovod, parallel data processing with Ray, Dask and Spark, and DAG-based pipelines with schedules and retries. Experiment tracking, dashboards, logs and evaluation are built in, and the same control plane also schedules agent executions and browser-based workloads. Deployments are self-hosted on any Kubernetes โ on-prem, in a VPC or on managed cloud clusters โ so sensitive data never leaves your infrastructure.
The open-source Community Edition is free with all core features. Polyaxon also offers a fully managed Cloud control plane and an on-premise Enterprise Edition with premium support, so the platform is listed as Freemium.
ML platform teams and research organisations that need reproducible, governed model training on their own Kubernetes infrastructure instead of a hosted black box, from single-GPU experiments to multi-node training fleets.