Matrixhub

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. By deploying MatrixHub on your own infrastructure, you eliminate dependency on public internet for model distribution, achieving speeds up to 25.8 GB/s on intranet and sub-second startup times. MatrixHub is designed for SREs and Algorithm Engineers managing massive model weights. It offers a transparent HF proxy: simply set HF_ENDPOINT to your MatrixHub instance, and all existing training/inference code works unchanged. On-demand caching pulls a model once and caches it locally, slashing redundant traffic and accelerating cluster-wide distribution. RBAC and audit logs provide fine-grained permissions, project-based isolation, and comprehensive trails for every upload/download. Storage-agnostic, MatrixHub supports local filesystems, NFS, and S3-compatible backends (MinIO, AWS) for unlimited model capacity. Key use cases include zero-wait distribution across 100+ GPU nodes, air-gapped delivery with integrity protection and malware scanning, private registry for fine-tuned weights with tag locking and CI/CD integration, and global multi-region sync for high availability. Deployable via Docker Compose or Helm, MatrixHub is free and open-source under Apache 2.0, making it the Harbor for models—a privacy-first, performance-optimized alternative to Hugging Face.