kubeflow/kubeflow is now a signpost: what the repository holds, and where Kubeflow itself is built

A repository with a README, an OWNERS file and no source code

The top level of kubeflow/kubeflow holds .github/, .gitignore, CONTRIBUTING.md, LICENSE, OWNERS, README.md, ROADMAP.md and a logo/ directory. Nothing else. There are no manifests, no Go or Python packages, no Helm chart, and GitHub reports no primary language for it.

The README is explicit about the role: this repository serves primarily as a gateway to Kubeflow subprojects and shared project metadata, and Kubeflow development happens in the individual subproject repositories. Cloning it gives you a logo, a roadmap and a governance file. A reader who arrives from a search for Kubeflow on GitHub and expects a codebase to audit or build will find nothing to compile.

Releases stopped at v1.10.0, then a release named redirect

The release history shows the shift. v1.9.2 was published on 2024-10-23 and v1.10.0 on 2025-03-25. The next entry, on 2026-04-29, is tagged redirect and titled Kubeflow Subprojects Releases. There has been no versioned Kubeflow release in this repository since v1.10.0.

The consequence is practical. Teams that subscribed to releases on kubeflow/kubeflow to learn about new Kubeflow versions will not see component releases here any more. Release notes, version tags and upgrade information now belong to each subproject. If your runbook says "check github.com/kubeflow/kubeflow/releases before upgrading", it points at a page that no longer carries that information.

The repository itself is still touched: the last push was on 2026-09-30.

How to install Kubeflow on Kubernetes does not start in this repository

The README has no installation section, no prerequisites and no commands. It does not state a minimum Kubernetes version, cluster size, or which cloud providers are supported. That is consistent with its gateway role, but it means a reader cannot plan an install from it.

The governance section names a Kubeflow Distribution Committee alongside the Outreach and Steering committees, which tells you distributions exist as a governed concept. The README does not list them or explain how they differ. The homepage, kubeflow.org, is the starting point the README links to for documentation, community and contribution guides.

For a platform team, the honest summary is that "installing Kubeflow" is a choice of which subprojects to run, made from documentation that lives outside this repository.

Four principles that the README asserts but does not back with numbers

The README frames Kubeflow with four principles. Simple: run workloads at any scale without becoming a Kubernetes expert. Portable: use the same code on a local laptop, on-premises, or in any cloud. Scalable: manage hyperscale training jobs and high-throughput AI agents. Composable: mix and match tools across the AI lifecycle.

These are goals, and the README gives no benchmarks, scale figures or compatibility tables to test them against. "Without becoming a Kubernetes expert" deserves particular scepticism from anyone who will operate the platform rather than use it. The workloads still run on Kubernetes, and the cluster still needs someone who understands it.

The composable principle is the one with direct consequences. Because Kubeflow is a set of separate projects, a team can adopt only the pieces it needs. The flip side is that the team also owns the integration between those pieces.

Kubernetes runs the containers, Kubeflow is the AI stack on top

Kubeflow does not replace Kubernetes. The README calls it the Kubernetes-native stack for data and AI workloads, built from modular open source projects, with a mission to bridge the data, AI and cloud native ecosystems and help teams move models, agents and AI applications into production. Kubernetes is the substrate; Kubeflow is tooling that runs on it.

That dependency is the main reason Kubeflow is the wrong tool for some teams. A data science group with no Kubernetes cluster and no one to run one takes on that whole layer before getting any Kubeflow benefit. MLflow, by contrast, is a single project that does not require Kubernetes to run, so a team whose need is experiment tracking rather than a cluster-wide AI platform has a lighter starting point there. The README itself makes no comparison with MLflow, Airflow or any other tool.

Working groups under three committees decide what ships

Kubeflow is described as community-led, maintained by Kubeflow Working Groups under the guidance of the Outreach Committee, the Distribution Committee and the Steering Committee. Links for each go to the governance page on kubeflow.org. The OWNERS file in the repository root lists who can approve changes here, and CONTRIBUTING.md and the contributing guide on kubeflow.org cover how to take part.

For an adopter, this structure answers a specific question: who do you talk to about a bug or a feature? Not this repository. Issues about a component belong in that component's repository, where its working group triages them. ROADMAP.md in this repository is the one place to read project-wide direction.

A relative architecture link that resolves nowhere on GitHub

The mission paragraph links "the AI lifecycle" to /docs/started/architecture/#kubeflow-landscape-in-the-ai-lifecycle. That is a site-relative path written for kubeflow.org. Rendered on GitHub, it resolves against the repository, which has no docs/ directory, so the link goes nowhere. Readers who want the architecture overview should open it on kubeflow.org instead.

The licence is Apache-2.0, which permits commercial use, modification and redistribution with notice and attribution requirements. That LICENSE file covers this repository. Since the actual software lives in the subproject repositories, check each subproject's own licence file before you redistribute or bundle it, rather than assuming the umbrella repository's terms carry over.

Editorial conclusion

Platform teams that run Kubernetes and want composable data and AI tooling should evaluate Kubeflow, but not from this repository: go to kubeflow.org and the individual subproject repositories, which is where code, issues and releases now live. Anyone who wants a single package with one version number to install should not expect it here, since the last versioned release in kubeflow/kubeflow was v1.10.0 on 2025-03-25. Before planning an install, confirm which subprojects you need and check each one's own repository and licence.

Frequently asked questions

Official sources

1. kubeflow/kubeflow on GitHub
2. License: Apache-2.0
3. Project website
4. README
5. Releases