Polyaxon

Polyaxon is an open source control plane for running machine learning and AI workloads on Kubernetes. It schedules, tracks and governs training jobs, services, notebooks, agent sessions and sandboxes on the customer's own clusters, so data, models, logs and artifacts stay inside infrastructure the customer controls. It is aimed at ML and platform teams that need one place to manage GPU training and inference. Capabilities include fine tuning and distributed training with PyTorch DDP, MPI and Horovod, Ray, Dask and Spark workloads, hyperparameter sweeps, DAG pipelines, experiment tracking and comparison, a model and artifact registry, queues with concurrency and quota limits, and organization level roles, audit logs and policies. Workloads are described in a Polyaxonfile written in YAML, JSON, Python, Go or TypeScript. The core is Apache 2.0 licensed and can be self hosted through a Community Edition Helm deployment, while a managed control plane connects to the customer's clusters. Published hybrid cloud pricing has three tiers. Platform is $555 per month and includes 3 developer seats, one compute cluster and 50 concurrent runs. Teams is $1500 per month and includes 6 developer seats, team spaces, an IP allow list and priority support. Enterprise is custom priced and adds SSO and SAML, custom branding and support SLAs. Extra developer seats cost $99 per month and read only seats $11 per month. Every paid tier offers a free trial, and the site notes a discount for academics. Self hosted pricing is presented separately on the same page. The site also announces an upcoming Polyaxon v3 release.