If you have ever run a Python script that used a Hugging Face model, used Whisper for transcription, or experimented with sentence embeddings, there is a good chance you have gigabytes of model files cached on your Mac that you have completely forgotten about.
Hugging Face does not ask before caching. It just does it, quietly, in a hidden folder.
The fast answer
Hugging Face keeps a local cache for every model it downloads. The cache lives in a hidden system directory. Each model gets its own folder, and those folders can easily reach 5 to 50 GB depending on what you downloaded.
This is by design — caching means you do not re-download models every time you use them. But it also means the cache grows without any automatic cleanup, and most users never think to look there until their disk is full.
Why it is not as simple as deleting a folder
The Hugging Face cache is structured around repositories, not individual files. When you download a model, you get the full repo: weights, config files, tokenizer files, and sometimes multiple model variants. The files are organised with symlinks pointing between a blobs directory and a snapshots directory.
If you delete only the symlinks, you free no space at all. If you delete only the blob files and leave the symlinks, you silently break every script that references that model. Doing this correctly means understanding the relationship between both directories — or using a tool that handles it for you.
Want to skip the hidden-folder hunt?
LLM Cleaner scans your Mac for local AI models, caches, indexes, and project memory — then shows what you can review, reveal, export, or safely move to Trash.
The shared cache problem
The Hugging Face cache is shared between every tool and script on your machine. Transformers, Diffusers, Sentence Transformers, Whisper, and any other library that uses Hugging Face models all write to the same location.
That means a model you downloaded for one project might be actively used by three others. Deleting it to reclaim space removes it everywhere — including from projects you are currently using.
Before you clean anything
The safe approach is to see the full picture first. Which repos are cached, how large each one is, when each was last accessed, and which ones are being used by active projects versus sitting from old experiments.
Without that picture, any cleanup is a guess.
Not sure what is safe to delete?
LLM Cleaner separates models, rebuildable caches, and project memory so you do not treat everything like junk.