Jozu Hub Quick Start

This guide walks you through Jozu Hub by example. You'll investigate a security issue in a ModelKit, look at what's inside the ModelKit package, import a clean model from Hugging Face, and run it on your laptop. It takes about 10 minutes.

Before you start

You'll need:

• A Jozu Hub account. Sign up if you don't have one.
• Docker installed locally if you want to run the model in Step 7 (this step is optional).

The flow uses Jozu Hub SaaS at jozu.ml. The same flow works on self-hosted Jozu Hub, with two differences noted along the way.

1. See AI risks across your organization

After signing in, the Status page is the first thing you see. It surfaces three things across every ModelKit in your registry: security issues, license issues, and (on self-hosted only) models, agents, and MCP servers currently running in production.

On self-hosted Jozu Hub, this is a single-pane view of risk across your AI supply chain - what's broken, what's licensed wrong, and what's actually live. On SaaS, the security and license panels work the same way and surface real findings across every ModelKit you can see; the production count is zero because SaaS Hub isn't connected to your runtime environments.

2. Investigate a compromised ModelKit

The Security Issues view lists ModelKits with critical and high severity findings from Jozu Hub's scanners. Find and click into the jozu-quickstarts/fraud-detection repository at tag 2026-05-04.

The Security Report tab shows what Hub flagged. Jozu Hub scans across nine threat categories that container scanners don't cover - malicious code in model weight files, backdoored weights, prompt injection patterns, license issues, and others. See Security Scanning for the full list and how scoring works.

Scan results aren't just displayed in the UI. Each result is a signed attestation bound to the ModelKit's digest, so downstream policy engines can refuse to load a model that doesn't meet your bar. For self-hosted Jozu Hub, this works even in air-gapped environments. See Attestations for how this works in practice.

3. See what's inside the ModelKit

Click the ModelKit Contents tab on the same repository. This shows how the layered packaging of KitOps allows you to package your AI project with all the context teams, agents, models, or security and compliance teams need.

For fraud-detection, you'll see:

• The model file itself (fraud-xgb serialized as a Pickle file)
• A SKILL.md giving context for the fake financial institution and intended use
• The datasets the model was trained on
• The original Jupyter notebook, MLFlow run script, and Python dependencies
• Docs covering licensing and key graphs from the MLFlow run that produced the trained model

Each section has a button in the top right that copies a kit unpack command extracting only that part of the ModelKit - useful when the model file is large but you only want the prompts or datasets.

The parameters portion of the Model section is free-form JSON. It's where teams capture training run IDs, evaluation metrics, hyperparameters, and pointers to upstream artifacts. Here we've captured key data from our MLFlow training session. This can inform deployment and placement for inference, and gives information that might be required by compliance teams.

4. Compare versions

Click the ModelKit Diff tab. The right-hand pane shows the tag you're currently viewing (2026-05-04). Click the tag name at the top of the left-hand pane and pick an earlier tag to compare against.

The diff shows what changed at the layer level: which files were added, removed, or modified, with size differences. This is the AI-artifact equivalent of git diff for a model package - a way to see what actually changed between two versions of a model, not just that the version string bumped. In many organizations there are different pipelines for model, agent, and dataset deployments - the diff can help DevOps teams know which need updating.

5. Import a clean model from Hugging Face

In the top navigation click Add Repository, then Import from Hugging Face.

Paste the Hugging Face URL https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF and click Import. No token needed - this model is open.

Hub will package the artifact as a ModelKit under your account, run the security scan, and email you when it's done. Imports for small models take a few minutes; large models can take longer.

For the rest of the tour we'll use a pre-imported copy at jozu-quickstarts/qwen2.5-0.5b-instruct-gguf so you don't have to wait for your own import to finish.

6. View the clean security report

Open jozu.ml/jozu-quickstarts/qwen2.5-0.5b-instruct-gguf and select the may-18 tag.

The Security Report tab shows no findings so the model is clean. As with the compromised model in Step 2, this scan result is a signed attestation attached to the ModelKit's digest. If you pull this ModelKit into an air-gapped environment, the attestation comes with it and verifies offline.

7. Run it locally

Click the Deploy tab. Jozu Hub readies an inference container directly from the ModelKit, with no image to build or pre-pull (learn more about Jozu's Rapid Inference Containers). Select Llama.cpp as the runtime (GGUF models are designed to run on Llama.cpp).

In a terminal with access to docker type:

docker run -it --rm \
  --publish 8000:8000 \
  jozu.ml/jozu-quickstarts/qwen2.5-0.5b-instruct-gguf/llama-cpp:may-18

Once the container is up, open http://localhost:8000/ in your browser. You're now talking to a Qwen 2.5 0.5B Instruct model running on your laptop, served from a container that Jozu Hub composed on the fly from the ModelKit and an approved base image.

The reference in the Deploy tab can also be used in a Kubernetes deployment YAML if you want to run the same model in a cluster. See Rapid Inference Containers for how RICs work and how to configure additional runtimes on self-hosted Jozu Hub.

Next steps

You've seen Jozu Hub investigate, curate, and deploy a ModelKit. From here:

• Push your own ModelKit - package a model from your environment and push it to Hub.
• Organizations - set up shared namespaces with role-based access (self-hosted only).
• Attestations - attach signed evaluation results, build provenance, and human-in-the-loop approvals to your ModelKits.
• Security Scanning - what Hub scans for, how scoring works, and how to gate deployment on scan results.
• Jozu Agent Guard - the runtime that enforces ArtifactPolicy, ToolPolicy, and GuardrailPolicy on the ModelKits you've curated in Hub.