// AI NATIVE STACK

AI Native › AI Native Infra › Model Asset and Registry › KitOps

CRASH COURSE · AI-NATIVE · intermediate · 9 min read · ModelKit

KitOps — package models like containers, with a ModelKit manifest.

model-registryai-nativekitopsocimodelpackaging

TL;DR — KitOps standardizes the delivery of AI assets by packaging a model, its code, dependencies, and metadata into a portable OCI artifact called a ModelKit. It turns "download weights from random storage" into "pull a versioned artifact from a registry".

What it is

KitOps is an open-source packaging standard and CLI for AI/ML workloads. A ModelKit is an OCI artifact that can contain model weights, prompts, code, parameters, and a declarative manifest. It lives in AI Native Infra › Model Asset and Registry.

Why it exists

Training artifacts are more than weights. You need a reproducible bundle that includes the model, tokenizer, config, preprocessing, and runtime assumptions. KitOps makes that bundle portable and registry-friendly, so teams can promote the same artifact from dev to staging to prod without hand-assembling files.

How it works

KitOps uses a ModelKit manifest to declare inputs and outputs. The CLI packages the referenced files into an OCI image layout and pushes it to a registry. Consumers pull the ModelKit, unpack it, and run it with the recorded metadata intact.

Key features

  • OCI-native — store models in any OCI registry.
  • Reproducible — model, config, and code travel together.
  • Registry agnostic — works with Harbor, ECR, GHCR, and on-prem registries.
  • Promotion friendly — tag ModelKits through environments like images.

Quick start

kitops init my-model
kitops pack --from ./bundle --tag registry.example.com/ml/my-model:1.0
kitops push registry.example.com/ml/my-model:1.0

When to use, when to skip

Use it when you want strict, registry-backed packaging for model delivery and promotion. Skip it if you only need an upstream model catalog like Hugging Face Hub or if your team hasn't standardized on OCI registries yet.

heads upKitOps is about packaging, not serving. You'll still need a deployment/runtime layer such as KServe, vLLM, or Kubernetes jobs.

vs / alongside

ToolRoleNote
KitOpsModel packaging in OCIPortable bundles
ORASGeneric OCI artifact push/pullLower-level primitive
Hugging Face HubHosted registryUpstream discovery
KServeServing layerConsumes packaged artifacts

References

Extra reads

Verified against KitOps docs, May 2026.

← AI Native Stack
© cvam — written in plaintext, served warm