A Container Registry
Is Not a Model Registry
Both use OCI standards. Only one was built for AI model security, governance, and deployment.
Docker Hub stores containers. It can hold AI models as OCI artifacts, but it has no model-specific metadata, security scanning, governance workflows, or deployment automation. Jozu Hub is a purpose-built model registry that uses the same OCI standard and adds what ML teams actually need: five-layer threat scanning, tamper-proof packaging with cryptographic attestations, audit trails, and Kubernetes deployment.
Docker Hub added AI support. Here's where it falls short.
Docker Hub recently introduced Model Runner and OCI Artifact support, making it technically possible to store and distribute AI models through Docker's infrastructure. Many ML teams use Docker Hub or private Docker registries as makeshift model registries because they already have the infrastructure in place.
It works until someone asks: "Was this model scanned before deployment?" or "Who approved this version for production?" or "Can we prove this model wasn't tampered with?" A generic container registry can't answer those questions. That's the gap Jozu fills.
Feature Comparison
| Capability | Jozu Hub | Docker Hub |
|---|---|---|
| OCI artifact storage | ✓ OCI 1.1 compliant | ✓ OCI artifact support |
| Model-specific metadata | ✓ Training params, metrics, lineage, framework | ✗ Basic JSON config only |
| AI security scanning | ✓ Five-layer: backdoors, poisoning, injection, adversarial, red teaming | ✗ Container vuln scanning only |
| Tamper-proof integrity | ✓ SHA attestation + signed provenance per model | ◉ Image digest verification (not model-specific) |
| Software bill of materials | ✓ Automatic SPDX 3 for every ModelKit | ✗ Not available for models |
| Model governance workflows | ✓ Approval stages, policy enforcement, deployment gates | ✗ No model lifecycle management |
| Audit trails | ✓ Per-version scan results, tamper-evident compliance logs | ✗ Basic access logs only |
| Model discovery & search | ✓ Browse, search, trending, favorites | ◉ Tag-based search only |
| Hugging Face import | ✓ One-click import with auto-scanning | ✗ Manual re-packaging required |
| Kubernetes deployment | ✓ Auto-generates manifests + inference containers | ✗ Pull only; deployment is DIY |
| Inference optimization | ✓ Rapid Inference Containers | ✗ No inference features |
| On-prem / air-gapped | ✓ Full behind-firewall deployment | ◉ Docker Registry (self-hosted) possible, no ML features |
| Regulatory compliance | ✓ EU AI Act, ISO 42001, NIST AI RMF | ✗ No AI compliance features |
| Experiment tracking | ◉ Via MLflow integration | ✗ Not available |
The core gap: Docker Hub treats models as generic blobs. Jozu treats models as first-class citizens with metadata, security posture, governance state, and deployment readiness tracked per version.
What breaks when you use Docker Hub for AI
No model-specific security scanning
Docker Hub's scanning (via Docker Scout) detects known container vulnerabilities: CVEs in OS packages, library versions, and configuration issues. It doesn't detect AI-specific threats like backdoors injected into model weights, data poisoning during training, or adversarial manipulation. For ML models, the threat surface is fundamentally different from containers.
No governance or approval workflows
When a data scientist pushes a model to Docker Hub, it's immediately available. There's no concept of "staging," "approved for production," or "pending security review." No one signs off. No audit trail records who approved what. For regulated industries, Docker Hub can't be your system of record for model governance.
No model metadata at the registry level
Docker Hub stores tags and digests. It doesn't store training parameters, performance metrics, dataset lineage, framework versions, or compliance status. Teams end up maintaining separate spreadsheets or wikis to track model metadata, which defeats the purpose of a centralized registry.
No deployment automation for ML
Docker Hub gives you docker pull. For ML inference, you still need to build a serving container, configure the runtime, write Kubernetes manifests, set up health checks, and handle GPU allocation. Jozu auto-generates all of this from the ModelKit.
OCI-native. ML-intelligent.
Jozu doesn't abandon OCI. ModelKits are OCI-compliant artifacts built on the OCI Image Manifest Specification. They work with any OCI registry. The difference is what Jozu adds on top.
Each ModelKit bundles model weights, datasets, code, configs, documentation, and metadata into a single immutable package. Every artifact gets a SHA256 digest. The package gets a signed provenance attestation and an SPDX 3 software bill of materials. When pushed to Jozu Hub, the model is automatically scanned across five layers of AI-specific threats before anyone can deploy it.
Because ModelKits are OCI-native, they inherit container infrastructure benefits (layer caching, deduplication, registry compatibility) while adding the ML-specific intelligence that generic registries lack. Your DevOps team works with familiar tools. Your security team gets the governance they need.
When to Choose Each
Docker Hub works if:
- → You're storing models for personal projects or prototypes
- → You don't need security scanning specific to AI threats
- → Model governance and audit trails aren't required
- → Your team handles deployment manually
Choose Jozu when:
- → Models go to production in regulated industries
- → You need AI-specific security scanning, not just container scanning
- → Compliance requires audit trails, SBOMs, and provenance
- → You want automated Kubernetes deployment from the registry
- → On-prem or air-gapped environments are a requirement
- → Multiple teams need governance workflows for model approvals
Docker Hub to Jozu. Same OCI standard. Straightforward move.
Models already stored as OCI artifacts in Docker Hub (or any OCI registry) can be pulled and re-pushed to Jozu Hub. The KitOps CLI handles the packaging: kit pack wraps your model into a ModelKit with metadata, and kit push sends it to Jozu Hub where it's automatically scanned and governed.
You don't have to abandon Docker Hub. Many teams keep Docker Hub for container images and use Jozu Hub specifically for AI models. Each registry stays purpose-built for what it does best.
Frequently Asked Questions
Can Jozu Hub work with my existing Docker registry?
Yes. ModelKits are OCI-compliant artifacts and work with any OCI registry: Docker Hub, ECR, GCR, Harbor, Artifactory, and others. Jozu Hub can serve as your primary model registry while your existing Docker registry handles container images.
What's the difference between Docker Scout scanning and Jozu's scanning?
Docker Scout detects container-level vulnerabilities (CVEs in OS packages and libraries). Jozu's five-layer scanning is purpose-built for AI: it detects backdoors in model weights, data poisoning, code injection, prompt handling issues, and adversarial attack vectors. They address fundamentally different threat surfaces.
Does Jozu support Docker's Model Runner?
Jozu uses the same OCI standard that Docker Model Runner builds on. ModelKits are OCI artifacts that can be stored in any OCI registry. Jozu adds the security, governance, and deployment layer on top of what OCI provides.
Is Docker Hub free? Is Jozu free?
Docker Hub offers a free tier with paid plans starting at $9/month. Jozu's open-source component (KitOps) is free as a CNCF Sandbox project. Jozu Hub offers a free trial with enterprise pricing for full on-prem and SaaS deployments.
Why not just use a private Docker registry for models?
You can. But a private Docker registry gives you storage without intelligence. You'll still need separate tools for model metadata tracking, security scanning, governance workflows, and deployment automation. Jozu Hub consolidates all of this in a single, purpose-built platform.
Your Models Deserve More
Than a Container Registry
Keep your Docker registry for containers. Let Jozu handle the models.