A self-hosted model library. Download Hugging Face models onto your own storage, pin the revision you want, verify the copy, and keep an eye out for updates. Includes an MCP server and scoped keys for agentic use, and multiple locations and user accounts support enterprise deployments with tiered storage.
Use cases
1.8 GB or 1.8 TB, size doesn't matter: Weighthouse runs each download in a managed worker. It supports resuming stopped or failed transfers and keeps partial downloads safe across restarts. The default transport streams with a small, fixed memory footprint; on high-bandwidth links you can opt into Hugging Face's Xet transport for that extra push.
resume / retries / restart-safe / Xet opt-in
Large model collections usually require different types of storage: keep your archives cold in object storage, keep a working set warm on NAS or SAN, and stage the models your inference engines are actively using onto fast NVMe. Weighthouse tracks every copy across locations, and moves safely them between locations.
cold / warm / hot / safe moves
Every managed copy has a manifest. When the upstream repository is modified, Weighthouse shows you which files changed and waits for your call — nothing updates itself, and upstream removals are reported, never automatically applied. The catalog shows the Hub and your library side by side, so you always know which revision you're serving. Track notes on each model, and share your notes with other users.
pinned revisions / manifests / drift detection / shared notes
Instead of giving an agent your account credentials, Weighthouse issues scoped keys. With a key, an agent can search the Hub, read model cards and GGUF metadata, queue downloads, and read the library — the same operations a person gets in the interface, over a stable API. Research and ops pipelines no longer need a browser, or your password.
scoped keys / stable API / Hub research / no browser
Enterprise
From a rented GPU to racks of inference servers, the install is the same: one container, your storage, your keys. The parts teams lean on most:
Roles, per-account libraries, notes, and preferences.
Model notes travel with the team, pinned or latest.
Your own weights, private by default, shareable with the team.
Move copies between n-tiers of cold, warm, and hot storage on demand.
Drift is reported, never applied.
For larger multi-user installations and enterprises (or hoarders with thousands of models)
Scoped keys and a stable API for research and ops.
Open source, permissive licensing. 'Nuff said.
No team yet?
The install is one container and a storage path. The interface is a search box, a library, and a downloads page. Easy defaults out of the box help get you started quickly.