Matrixhub

MatrixHub is an open source, self hosted registry for AI models, built for enterprise inference at large scale. It is positioned as a drop in private replacement for Hugging Face, and the site compares its relationship to Hugging Face with the one between Harbor and Docker Hub. It is designed to speed up vLLM and SGLang workloads and runs alongside Kubernetes clusters. The core idea is to stop depending on the public internet for model weights. A transparent Hugging Face proxy means teams point their HF_ENDPOINT setting at MatrixHub and keep their existing training and inference code unchanged. On demand caching pulls a public model once and serves it to the whole cluster, which reduces repeated downloads and speeds distribution across many GPU nodes. The site cites intranet download speeds of 25.8 GB per second and speeds above 10Gbps across more than 100 GPU nodes at once. Role based access control, project isolation and audit logs cover every upload and download. Storage is agnostic and works with local filesystems, NFS and S3 compatible backends such as MinIO and AWS. Listed use cases are zero wait distribution, air gapped delivery with integrity protection and malware scanning, a private registry for fine tuned weights with tag locking and CI/CD integration, and global multi region sync with resumable replication between data centers. The audience is site reliability engineers and algorithm engineers managing large model weights. Deployment is through Docker Compose or Helm, with a quick start guide and a GitHub repository. The site states that it is open source and free for the community, and the home page carries a Coming Soon label. No paid plans or prices are published.