Use cases / Storage

Your model library doesn’t have to fit on one disk.

Large model collections take up space. You might keep the models you’re using on a fast SSD, with larger or less frequently used models on another drive or a network share. Once they’re spread across a few places, keeping track of what lives where gets harder.

Weighthouse brings those locations into one library. Configure named storage locations on local disks or mounted network filesystems, see where your models are stored, and move managed copies between locations as your needs change. Each copy’s manifest moves with its files, keeping its recorded revision and file selection with it.

Use the storage you already have

A storage location is just a directory that Weighthouse can access. It can live on a local SSD, a larger hard drive, or a mounted network share. Configure those locations on your server and give them names that make sense to you, like “fast” or “archive”.

Your models appear in one single library, regardless of their physical location. While browsing your library, it’s easy to see where each model is stored, and you can even filter by location to see what’s stored in each place.

You no longer need to browse every drive where you keep your models, or keep a separate inventory somewhere that quickly gets out of date.

Move models, don’t redownload them

Fast storage tends to fill up quickly. If a model isn’t part of your current workload, you can move its managed copy to a larger drive or a network share. You can always bring it back later when you need it again.

Moving models is easy. Just choose a destination from the model page and Weighthouse moves the files along with their manifest and records their new location in the library.

When moving between filesystems, Weighthouse first copies the files to the destination, checks the file count and total size, and only removes the source copy if those match.

Your models are still ordinary files

Weighthouse stores your downloads in ordinary directories, preserving the repository’s original file layout. Your weights, configuration, tokenizer files, and documentation stay fully accessible to the tools you already use.

There’s no export step before you can use a model. Just point your tool or engine at the repository folder or weight file it expects. Weighthouse adds a manifest alongside the model files to keep track of the managed copy.

The library will also read and index models that you already downloaded, from before you began using Weighthouse. Configure the directories containing those model folders as storage locations and scan them into Weighthouse. It indexes those files without modifying them.

Keep models in S3-compatible storage

If you use object storage, Weighthouse can keep your managed models in Amazon S3 or a compatible service such as MinIO, Ceph, or Cloudflare R2. With that backend enabled, completed downloads are synced to object storage, and your model directory acts as a local working cache.

Once a complete copy is stored in S3, you can remove the local cache to free up disk space. The model stays in your library, and you can restore its files when you need them again.

Your inference tools still need filesystem access to the model files, so restore a model to the local cache before loading it.

Read the storage documentation · Managed downloads · Revision tracking