# Olah

> A self-hosted lightweight Hugging Face mirror service that caches models, datasets, and spaces at the file block level to speed up downloads and save bandwidth.

- **Category:** AI
- **Licence:** MIT
- **Latest version:** v0.5.4
- **Open source:** oss
- **Languages:** Python
- **Platforms:** Docker, Docker Compose, pip, Build from Source
- **Tags:** Proxy, Datasets, LLM, AI, Caching, Self Hosted, Self Hosting, Huggingface, Spaces, Mirror, Models, Mirror Server

## Description

Olah is a self-hosted lightweight Hugging Face mirror service that implements mirroring at the file block level rather than just reverse proxying. It caches files in chunks as users download them, so subsequent downloads are served from the local cache. It supports models, datasets, and spaces mirroring with configurable access control and cache policies including LRU, FIFO, and size-based eviction. Can be installed via pip, from source, or Docker. Runs on port 8090 by default and clients connect by setting HF_ENDPOINT to the mirror URL.

## Links

- [Homepage](https://github.com/vtuber-plan/olah)
- [Docker](xiahan2019/olah)
- [GitHub Repository](https://github.com/vtuber-plan/olah)
- [Documentation (English)](https://github.com/vtuber-plan/olah/tree/main/docs/en)
- [Documentation (Chinese)](https://github.com/vtuber-plan/olah/tree/main/docs/zh)

---
Source: https://selfhost.directory/project/olah