finds.dev

kitops-ml/kitops

An open source DevOps tool from the CNCF for packaging and versioning AI/ML models, datasets, code, and configuration into an OCI Artifact.

View on GitHubHomepage ↗

★ 1,372GoApache-2.0updated Jun 2026

// the take

KitOps packages AI/ML models, datasets, code, and configs into OCI artifacts stored in standard container registries — think 'Docker for ML assets.' It's a CNCF sandbox project targeting platform engineers and MLOps teams who need reproducible, signable, auditable model artifacts without standing up a separate model registry. Most useful in regulated environments or anywhere the question 'exactly which weights went to prod?' matters.

OCI reuse is the right call — you get content-addressable storage, layer deduplication, and compatibility with every enterprise registry (Harbor, ECR, Artifactory) for free, with no new infrastructure to operate. Selective layer unpacking means a downstream service pulling only the model weights doesn't have to download the full dataset bundle, which matters when weights are 10GB+. Cosign integration plus SHA-256 digests gives you a real provenance chain that compliance frameworks like EU AI Act actually ask for, not just a README checkbox. The CNCF ModelPack spec backing means the format has vendor buy-in (Red Hat, PayPal, ByteDance), reducing the risk that this becomes a dead format in two years.

The Kitfile is one more YAML config file in a project that already has Dockerfiles, Helm charts, and pipeline definitions — teams will feel the friction of keeping it in sync, and there's no obvious mechanism to auto-generate it from training pipeline outputs without glue code. The 'dev mode' (local inference server wrapping llama.cpp) is bolted on and feels like a separate product; it muddies the positioning and the frontend Vue app ships inside what should be a CLI tool. Stars (~1.4k) and fork depth are low for a project making 'enterprise standard' claims, meaning community-driven bug fixes and integrations will be slow compared to established tools like DVC or MLflow. Documentation covers the happy path well but is thin on failure modes — what happens when a push partially fails mid-upload to a registry that doesn't support chunked OCI manifests?

We read the README, the directory tree and a key source file in June 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
1,372
Forks
172
Language
Go
License
Apache-2.0
Last commit
11 Jun 2026

// nearby

If kitops isn't quite it.

All repos →