Kubeflow
CNCF-graduated open-source AI platform on Kubernetes: pipelines, distributed training, tuning, notebooks and model hub
Overview
Kubeflow is "the open source AI platform" for Kubernetes: a family of projects you can install together or one at a time. The main parts are Kubeflow Pipelines (ML workflows), Trainer (distributed training), Katib (hyperparameter tuning), Notebooks, Hub (formerly Model Registry), the Spark Operator and a central Dashboard. The Cloud Native Computing Foundation announced Kubeflow's graduation on August 17, 2026, after incubation since 2023. The bundled install is now the Kubeflow Community Distribution on a twice-yearly calendar: 26.10.0 shipped on November 6, 2026 and retired Notebooks v1, TensorBoard and PVCViewer. 2026 also brought Trainer v2.4, Kale 2.0 and a Kubeflow MCP server. Everything is Apache-2.0; the kubeflow/kubeflow repository is now a gateway, with development in the subproject repos.
The Verdict
Who Should Use Kubeflow?
Best For
- Platform teams that already run Kubernetes and want a vendor-neutral ML stack
- Organizations that need on-prem or multi-cloud portability
- Teams that only need one piece (for example Pipelines or Trainer) on Kubernetes
Not Ideal For
- Small data science teams without Kubernetes skills
- Anyone who wants a hosted product with a support contract out of the box
- Quick experiment tracking (MLflow or W&B are lighter)
What's Great
- Free, portable and Kubernetes-native; runs wherever Kubernetes runs
- CNCF graduated, with a security audit and formal governance behind it
- Modular: adopt Pipelines, Trainer or Katib without the whole platform
- G2 reviewers praise the pipeline design and scale
Watch Out For
- Complex setup and a steep learning curve
- Ongoing ops burden to upgrade and secure the platform
- Users in the 2026 SDK survey name infrastructure complexity, hard debugging and slow dev cycles
- Component retirements (Notebooks v1, TensorBoard) mean migration work
Pricing
View all features & details
Components
- Kubeflow Pipelines
- Trainer (distributed training)
- Katib (hyperparameter tuning)
- Notebooks
- Hub (formerly Model Registry)
- Spark Operator and Dashboard
2026 Changes
- CNCF graduation (August)
- Community Distribution 26.03 and 26.10
- Trainer v2.4, Kale 2.0, Kubeflow MCP
- Notebooks v1, TensorBoard, PVCViewer retired in 26.10
Real-World Usage
Community
- CNCF graduated project (TOC vote July 2026)
- Two distribution releases a year, timed before KubeCon EU and North America
How Teams Use It
- Mainly as the orchestration layer for ML on Kubernetes, per the 2026 SDK survey
- 2026 blog posts cover Trainer v2.4, Kale 2.0 and agent access via Kubeflow MCP
How It Compares
| Feature | Kubeflow | Metaflow | ZenML | Vertex AI |
|---|---|---|---|---|
| Model | Open-source platform on Kubernetes | Open-source Python framework | Open-source pipelines plus Pro SaaS | Managed Google Cloud service |
| Runs on | Any Kubernetes | AWS, Azure, GCP, Kubernetes | Any stack | Google Cloud |
| Training and tuning | Trainer, Katib | Per-step cloud compute | Via integrations | Managed |
| Ops effort | High | Low to medium | Medium | Low |
Sources: kubeflow.org · GitHub · Kubeflow blog · Kubeflow 1.11 notes
Is this your tool? Add a badge