The Three Problems Modelpack Was Built to Solve

▶ Watch (0:02)

AI and ML teams face three concrete storage and distribution problems. First, model files arrive as loose archives, folders, and mixed data types with no standard packaging format. Second, those models often carry proprietary weights and metadata that need access control. Third, each serving framework, whether Python-based or Kubernetes-hosted, handles content differently. Modelpack addresses all three by wrapping models in a single, vendor-neutral standard rather than leaving each team to invent their own approach.

Modelpack as a CNCF Sandbox Project

▶ Watch (1:06)

Modelpack entered the CNCF sandbox in May 2025. The project defines a vendor-neutral, open-standards method for managing AI and ML models. It packages models as OCI artifacts, the same artifact format already used to distribute Helm charts. Teams already inside the cloud native ecosystem get a familiar distribution mechanism. No proprietary model store required.

Reusing a Decade of OCI Infrastructure

▶ Watch (1:36)

OCI has more than a decade of tooling for signing and distributing content. Modelpack inherits all of it. Teams can sign model artifacts with the same tools they use to sign container images. SBOM generation works through existing tooling as well. CI/CD integration with Jenkins or GitHub Actions requires no new connectors. Any team already running container registries gets version management, access control, and security compliance for their models at no extra infrastructure cost.

Publishing and Consuming Models with the Modelpack CLI

▶ Watch (3:50)

Publishing a model takes three steps: point Modelpack at a model repository, use the CLI to build a model artifact, then push it to a container registry. Block asked the audience how many already run a container registry, and the answer determined eligibility: if you have one, Modelpack works today. At runtime, the modelctl command pulls the artifact locally, or teams can use Docker, Podman, the Kubernetes CSI driver, or the image volume source to consume model content directly in cluster.

Notable Quotes

Model Pack is a vendor neutral open standards way for managing AI and ML models. Andrew Block · ▶ 01:08

You have security built in because guess what? You’re using OCI content. We have over a decade’s worth of of technologies and and and tooling around OCI for how you potentially sign and distribute content. Andrew Block · ▶ 03:14

How many of you have a container registry? Raise your hand. You can go ahead and use Model Pack then. Andrew Block · ▶ 04:03

Key Takeaways

  • Modelpack joined the CNCF sandbox in May 2025 as a vendor-neutral AI model packaging standard.
  • Models packaged as OCI artifacts inherit signing, SBOM, and registry tooling with no new infrastructure.
  • Runtime consumption works via modelctl, Docker, Podman, or the Kubernetes CSI driver.

About the Speaker(s)

Andrew Block is a Distinguished Architect at Red Hat. He works with organizations adopting open source solutions, with focus on cloud native, security, and emerging technologies. Block has authored multiple technical publications and speaks frequently on these topics.