matrixhub

matrixhub-aiApache-2.0活跃维护

自托管的开源 AI 模型仓库，兼容 Hugging Face，并针对 vLLM/SGLang 推理做加速优化，方便在自有集群里统一管理和分发模型。

An Open-source, self-hosted AI model hub with Hugging Face compatibility, accelerating vLLM/SGLang performance.

v0.2.0发布于 2026-09-18 13:41 UTC366star47fork同步于 2026-11-10 18:01 UTC

模型仓库自托管推理加速模型分发 Kubernetes

对比 GitHub 官网 artificial-intelligence dynamo huggingface kubernetes llm

怎么上手

由 README 提炼，执行前请确认

原文

README 未给出具体命令，仅指向 Getting Started 与 Installation 文档（Docker Compose / Helm）

硬件要求按 README 说明

README 未说明

有 Docker 镜像

下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼

没人工点评？写下你的经验

自托管 AI 模型仓库，兼容 Hugging Face，为 vLLM/SGLang 推理加速

自托管的开源 AI 模型仓库，兼容 Hugging Face，并针对 vLLM/SGLang 推理做加速优化，方便在自有集群里统一管理和分发模型。

核心能力

作为 Hugging Face 的私有替代，可直接替换使用
面向大规模企业推理场景的模型注册中心
加速 vLLM 与 SGLang 的模型加载与分发
提供 Docker Compose 与 Helm 两种部署方式
提供在线 Demo，可用 admin/changeme 直接试用

容易踩的坑

README 未给出安装命令，需先查阅官方 Getting Started 与 Installation 文档；Demo 账号 admin/changeme 仅供评估，可能随时被重置

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

3 个版本v0.2.0

全部 release

v0.2.0Release v0.2.0最新正式版2026-09-18 · 2026-09-18 13:41 UTC
v0.2.0

Changes

Feature
- Added SSH Git access for model repositories with public-key authentication, automatic model registration during clone and push, and configurable SSH port exposure for Docker Compose and Helm deployments. (#697, @setsunakute)
- The registry management UI now supports creating and managing MatrixHub-type registries. (#710, @lsq645599166)
- Prepare config.yaml in the same folder with docker-compose.yml to make docker compose setup steps easier (#814, @yitingdc)
- Refactor(chart): use StatefulSet for MySQL with volumeClaimTemplates (#890, @wzshiming)
- Guide users to create a registry when the target registry selector is empty during project creation. (#957, @linghaoSu)
- Added SQLite support for single-node deployments. Docker Compose now uses SQLite by default, while MySQL remains available through an explicit Compose configuration. Helm/Kubernetes deployments do not support SQLite. (#962, @setsunakute)
- Improve guidance for the Registry Base URL field when creating a registry: provider-specific hints and defaults, a confirmation before overwriting a customized URL on provider switch, and a warning for addresses that look like a specific repository. (#995, @linghaoSu)
Bug Fixes
- Copy buttons now work when MatrixHub is served over HTTP. (#678, @lsq645599166)
- Robot accounts now enforce disabled and expired states and all-project permissions, while model commits made with robot credentials show the correct uploader. (#688, @setsunakute)
- Long registry names now wrap correctly in the delete confirmation dialog. (#718, @lsq645599166)
- Project settings now display a success notification after changes are saved. (#719, @lsq645599166)
- Creating a model now rejects names longer than 251 characters. (#729, @scydas)
- Model parameter counts are now displayed correctly for single-file safetensors models and sharded models without total-size metadata. (#739, @scydas)
- Fixed model size display to use repository tree content size instead of Git repository disk usage. (#759, @scydas)
- Fix Markdown previews to consistently use the light color scheme. (#761, @vivian-xu)
- Creating a model now rejects invalid names and normalizes model identifiers to avoid detail-page 404s caused by leading or trailing whitespace. (#781, @scydas)
- Update
v0.1.1Release v0.1.12026-07-17 · 2026-07-17 04:15 UTC
查看更新说明
v0.1.1

Welcome to the v0.1.1 release of MatrixHub! Compared with version:v0.1.0, version:v0.1.1 has the following updates.
- Added deploy/config.yaml beside deploy/docker-compose.yml so Docker Compose setups can curl both artifacts from the same release tag and start with fewer manual config steps.
v0.1.0Release v0.1.02026-06-30 · 2026-06-30 06:21 UTC
查看更新说明
v0.1.0

MatrixHub v0.1.0 is the first official release. It delivers most features in Phase 1: Key Workflows and Enterprise Feature Baseline described in the roadmap: a self-hosted, Hugging Face–compatible model registry for private inference clusters.

Downloads

Container image
```
ghcr.io/matrixhub-ai/matrixhub:v0.1.0
```
Images are published for linux/amd64 and linux/arm64, signed with Cosign (keyless), and ship with an SPDX SBOM attachment.

Verify signature (optional):
```
cosign verify ghcr.io/matrixhub-ai/matrixhub:v0.1.0 \
  --certificate-identity-regexp='.*' \
  --certificate-oidc-issuer=https://token.actions.githubusercontent.com
```
Helm chart (OCI)
```
export CHART_VERSION=0.1.0
export NAMESPACE=matrixhub

helm install matrixhub oci://ghcr.io/matrixhub-ai/matrixhub \
  --version "${CHART_VERSION}" \
  --namespace "${NAMESPACE}" \
  --create-namespace
```
See the README and documentation site for Docker Compose, storage, and upgrade guidance.

Changelog since project inception

Feature
- Added a Hugging Face–compatible Git, Git LFS, and HTTP surface so clients such as vLLM and SGLang can use MatrixHub as a private model hub with HF_ENDPOINT.
- Added model repository lifecycle: create, browse, upload, download, and delete with project-scoped isolation.
- Added a Web UI for repository administration, user profile, access tokens, and platform settings.
- Added remote registry management with proxy-cache mode to localize public Hugging Face models on first request (pull-once, serve-all within the cluster).
- Added sync policies and an in-process job server to schedule registry replication tasks with resumable transfer foundations.
- Added token-based authentication, projects, roles, robots, and permission-aware project listing for multi-tenant operation.
- Added Docker Compose and Helm deployment paths with multi-arch container images published to ghcr.io.
- Added release automation: OCI Helm charts, Cosign signing, and SPDX SBOM generation for official tags.
- Added documentation site content at matrixhub.ai and CNCF Slack community channel (#matrixhub).
Documentation
- Published installation, operations, and development guides on the documentation site.
- Added project governance documents: Contributing, Security, Governance, and Maintainers.
Known limitations

Track follow-up work in GitHub issues.

用户点评

登录用户的真实使用记录，可按版本和硬件筛选

共 0 条

还没有人点评。第一个说出真实体验的人最有价值。

README 原文

未翻译

在 GitHub 打开

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.

For target users, use cases, and current capabilities, see the Overview.

🌐 Live Demo

Try MatrixHub instantly at demo.matrixhub.ai — no setup required.

Username	Password
`admin`	`changeme`

The demo is for evaluation only and may be reset at any time.

🚀 Quick Start

Follow the Getting Started guide to deploy MatrixHub and cache your first model.

For Docker Compose and Helm commands, see Installation.

For vLLM, SGLang, llm-d, Dynamo, and model distribution examples, see the Integrations guide.

Contributing and Development

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.

See Technical Design Docs for architecture and feature designs.

Community and Support

Join us in GitHub Discussions or the CNCF Slack #matrixhub for questions, ideas, and support.

Security

Please report vulnerabilities according to SECURITY.md.

License

MatrixHub is licensed under the Apache License 2.0.

怎么上手

项目说明

核心能力

容易踩的坑

最近发版

v0.2.0

Changes

Feature

Bug Fixes

v0.1.1

v0.1.0

Downloads

Container image

Helm chart (OCI)

Changelog since project inception