MatrixHub is an open-source, self-hosted AI model registry engineered for large-scale enterprise inference. It serves as a drop-in private replacement for Hugging Face, purpose-built to accelerate vLLM and SGLang workloads.
MatrixHub streamlines the transition from public model hubs to production-grade infrastructure:
- Zero-Wait Distribution: Eliminate bandwidth bottlenecks with a "Pull-once, serve-all" cache, enabling 10Gbps+ speeds across 100+ GPU nodes simultaneously.
- Air-Gapped Delivery: Securely ferry models into isolated networks while maintaining a native
HF_ENDPOINTexperience for researchers—no internet required. - Private AI model Registry: Centralize fine-tuned weights with Tag locking and CI/CD integration to guarantee absolute consistency from development to production.
- Global Multi-Region Sync: Automate asynchronous, resumable replication between data centers for high availability and low-latency local access.
- Transparent HF Proxy: Switch to private hosting with zero code changes by simply redirecting your endpoint.
- On-Demand Caching: Automatically localizes public models upon the first request to slash redundant traffic.
- Inference Native: Native support for P2P distribution, OCI artifacts, and NetLoader for direct-to-GPU weight streaming.
- RBAC & Multi-Tenancy: Project-based isolation with granular permissions and seamless LDAP/SSO integration.
- Audit & Compliance: Full traceability with comprehensive logs for every upload, download, and configuration change.
- Integrity Protection: Built-in malware scanning and content signing to ensure models remain untampered.
- Flexible Storage: Supports PVC-backed local file system and NFS storage today, with S3-compatible object storage planned for a future release.
- Reliable Replication: Policy-driven, chunked transfers ensure data consistency even over unstable global networks.
- Cloud-Native Design: Optimized for Kubernetes with official Helm charts and horizontal scaling capabilities.
Try MatrixHub instantly at demo.matrixhub.ai — no setup required.
Sign in with the public demo credentials:
| Username | Password |
|---|---|
admin |
changeme |
The demo is for evaluation only and may be reset at any time.
Download the Docker Compose files for a released version and start MatrixHub:
export MATRIXHUB_VERSION=v0.1.1
mkdir -p matrixhub && cd matrixhub
curl -fL \
"https://raw.githubusercontent.com/matrixhub-ai/matrixhub/$MATRIXHUB_VERSION/deploy/docker-compose.yml" \
-o docker-compose.yml
curl -fL \
"https://raw.githubusercontent.com/matrixhub-ai/matrixhub/$MATRIXHUB_VERSION/deploy/config.yaml" \
-o config.yaml
MATRIXHUB_IMAGE_TAG="$MATRIXHUB_VERSION" docker compose up -dFor a newer stable release, replace v0.1.1 with the version you want to run.
If port 3001 is already in use, set MATRIXHUB_HTTP_PORT before starting the stack, for example MATRIXHUB_HTTP_PORT=3002.
Open the MatrixHub web console:
http://127.0.0.1:3001
Sign in with the default local credentials:
| Username | Password |
|---|---|
admin |
changeme |
Change the default password before exposing the instance outside your local machine.
To stop the local stack:
docker compose downCurrently, the Helm chart supports PVC-backed storage for MatrixHub data. S3-compatible object storage is planned for a future release.
Make sure your cluster has a default StorageClass (kubectl get storageclass), or explicit storage settings for the PVCs this chart creates. For development clusters without a StorageClass, see development-only local storage setup.
MatrixHub publishes its Helm chart to GitHub Container Registry (ghcr.io) as an OCI artifact.
For a newer stable release, replace 0.1.1 with the chart version you want to run:
export CHART_VERSION=0.1.1
export NAMESPACE=matrixhubInstall the chart and expose the service via NodePort:
helm install matrixhub oci://ghcr.io/matrixhub-ai/matrixhub \
--version "$CHART_VERSION" \
--namespace "$NAMESPACE" --create-namespace \
--set apiserver.service.type=NodePortThe default installation uses the cluster's default StorageClass. The MatrixHub data PVC defaults to 50Gi, and the built-in MySQL PVC defaults to 8Gi. To change PVC sizes, add --set apiserver.storage.pvc.size=100Gi or --set mysql.persistence.size=20Gi to the command.
For other storage classes, existing PVCs, and other Helm settings, see the Helm chart README.
With the NodePort installation above, open:
http://<node-ip>:30001
Find a node IP with:
kubectl get nodes -o widehelm uninstall matrixhub --namespace "$NAMESPACE"This removes resources including the default PVCs created by the chart. To preserve data, use an existing PVC for MatrixHub data and an external database.
We welcome contributions.
See CONTRIBUTING.md for issues, pull requests, reviews, DCO, and release note requirements.
See Development Guide for local setup, testing, and generated code.
Join us in GitHub Discussions
or the CNCF Slack #matrixhub
for questions, ideas, and support.
Please report vulnerabilities according to SECURITY.md.
MatrixHub is licensed under the Apache License 2.0.