ORAS — push and pull AI artifacts with the OCI registry protocol.

TL;DR — ORAS (OCI Registry As Storage) is a generic CLI and spec for storing any artifact in an OCI registry. For AI, that means models, tokenizers, prompts, datasets, and metadata can live in the same registry that already stores containers. It turns registries into a universal artifact backend.

What it is

ORAS extends the OCI distribution spec beyond container images. It lets you push arbitrary blobs plus manifest metadata to any OCI-compliant registry. In the AI Native landscape it sits in AI Native Infra › Model Asset and Registry.

Why it exists

Registries already solve auth, replication, caching, and lifecycle. ORAS lets you reuse that machinery for ML artifacts instead of inventing a separate storage system. If your platform team already runs Harbor, ECR, or GHCR, ORAS gives you model storage with the same operational model as containers.

How it works

ORAS wraps files in an OCI manifest, attaches a media type, and pushes it to a registry. A pull retrieves the manifest and blobs, then the client reconstructs the artifact locally. Because it uses standard registry APIs, it benefits from auth, mirrors, and replication without extra services.

Key features

• Registry-native — works with any OCI registry.
• Artifact-agnostic — models, configs, prompts, datasets, anything.
• Simple CLI — easy push/pull without custom servers.
• Good for air-gapped infra — mirrors well inside private registries.

Quick start

oras login registry.example.com
oras push registry.example.com/ml/my-model:1.0 model.bin:application/octet-stream config.json:application/json

When to use, when to skip

Use it when you want model artifacts in your existing OCI registry and don't need a full model-catalog UI. Skip it if you want hosted discovery and community browsing like Hugging Face Hub.

vs / alongside

References

• ORAS — project home.
• ORAS docs — CLI and APIs.
• oras-project/oras — source.

Extra reads

• How-to guides — push/pull workflows.

Verified against ORAS docs, May 2026.