matrixhub-ai/matrixhub

What it solves

MatrixHub is a self-hosted AI model registry designed for enterprise-scale inference. It eliminates bandwidth bottlenecks when distributing models to many GPU nodes, provides a secure way to move models into air-gapped networks, and centralizes the management of fine-tuned weights to ensure consistency across development and production environments.

How it works

It acts as a private, drop-in replacement for Hugging Face, allowing users to redirect their HF_ENDPOINT to a private instance. It uses a "pull-once, serve-all" caching mechanism to accelerate distribution and supports P2P distribution and direct-to-GPU weight streaming via NetLoader. For enterprise governance, it includes Role-Based Access Control (RBAC), LDAP/SSO integration, and multi-tenancy.

Who it’s for

Infrastructure engineers and ML researchers in large-scale enterprises who need to host their own model registries for security, speed, and governance over their GPU clusters.

Highlights

• High-Performance Distribution: Supports 10Gbps+ speeds across 100+ GPU nodes and native P2P distribution.
• Transparent Proxy: Works as a transparent Hugging Face proxy with zero code changes required.
• Enterprise Security: Features RBAC, multi-tenancy, audit logs, and built-in malware scanning.
• Cloud-Native: Optimized for Kubernetes with official Helm charts and horizontal scaling.
• Global Sync: Automates asynchronous, resumable replication between different data centers.