KitOps v1.0.0 is Now Generally Available, Featuring Hugging Face to ModelKit Import

What is KitOps?

KitOps is a project, started in early 2024, that was inspired by the idea that we could define a better way of storing, sharing, and deploying AI/ML models. By using a structure related to the ubiquitous Docker container format, we gain many of the useful features of containers (such as immutability and simple distribution) while tailoring our implementation to be simple and easy to use. With a few commands, you can take a locally stored model, package it into the ModelKit format, and push it to most image registries currently used for sharing containers.

We've been working on the project for the past year and are proud to announce the release of KitOps v1.0.0. Here are some of the highlights from our first year:

Dev Mode

As large-language models become increasingly powerful and size efficient, running models locally is becoming a more and more common part of workflows. Using the dev command for the Kit CLI, many models can be started locally for inference without any additional setup.

PyKitOps SDK

We've written KitOps to be a portable CLI that can easily be run both locally and in CI/CD systems. However, we are also aware that most work in AI/ML takes place in a Python environment. To smooth the process, we designed a Python library that can be used to package ModelKits without having to switch away from your current Jupyter notebook.

Link: https://kitops.org/docs/pykitops/

Note: Our team is hard at working extending the KitOps documentation to fully cover this feature.

CI/CD support

We continue to work to ensure it's easy to integrate KitOps into existing flows. As a result, we've built modules for a number of common CI/CD tools:

• Dagger: Use composable Kit commands in your dagger pipelines
• MlFlow: Use Jozu Hub to serve as an Artifact Repository for MLFlow

We're continuing to look for more tools to integrate with. If there's an environment you would love to use Kit in, let us know!

New in v1.0: Import huggingface models to ModelKits directly

As we cross the v1.0 milestone in the project, we're proud to announce that we've made getting started with ModelKits even easier: the kit CLI can now import repositories from huggingface directly. Using the kit import command, you can take any model available on huggingface and convert it into a ModelKit that you can push to image registries such as DockerHub.

When you run, for example, kit import microsoft/phi-4, the Kit CLI will:

1. Download the microsoft/phi-4 huggingface repository locally
2. Generate the configuration needed to package the files in that repository into a ModelKit
3. Package the repository into a locally-stored ModelKit

Once this is done, you can simply kit push microsoft/phi-4:latest docker.io/my-organization/phi-4:latest and share it with your collaborators.