Is it safe to delete ~/.cache/huggingface?
- Yes, but pick what goes. It’s where Hugging Face’s Python libraries keep every model and dataset file you’ve downloaded, and they don’t delete any of it themselves. Use
hf cache rmandhf cache prunerather than deleting folders by hand. - What it costs you: a fresh download the next time your code loads a model or dataset you removed. Your Hugging Face login token sits in the same folder, so don’t delete the folder whole.
- Does it come back? Yes: when code asks for a model or dataset that isn’t cached, it downloads it again.
- SiftMac’s verdict: You decide
1. What it is
transformers, diffusers, datasets and other libraries built on huggingface_hub download files from the Hugging Face Hub into one shared cache, so a model fetched by one script is reused by the next. Hugging Face’s documentation is plain about the catch: cached files are never deleted, and when you download a newer revision of a model, the previous files are kept in case you need them again.
hub/: models, datasets and Spaces from the Hub, one folder each, named likemodels--org--name.datasets/: Arrow files thedatasetslibrary prepared from those downloads.xet/: data the Xet storage system keeps to speed up uploads.token: your User Access Token, if you’ve signed in.
2. Where it is
~/.cache/huggingface
That’s the default. HF_HOME moves the whole thing, HF_HUB_CACHE moves only hub, and HF_DATASETS_CACHE only the datasets cache. Inside each repository’s folder, blobs holds the actual files, snapshots has one folder of links per revision, and refs records which revision a branch such as main points to.
3. How big it gets
Model weights make up nearly all of it, and every revision you’ve downloaded counts. Hugging Face’s CLI lists each repository with its size and when you last used it, and can show revisions you haven’t touched in a while:
hf cache ls
hf cache ls --revisions --filter "accessed>30d"
du -sh ~/.cache/huggingface/*
Older installs without the hf command have huggingface-cli scan-cache for the same report.
4. How to delete it safely
Remove whole models or datasets you no longer use. hf cache rm asks before deleting, and --dry-run shows what would go:
hf cache rm model/<org>/<name> --dry-run
hf cache rm model/<org>/<name>
Then clear the leftovers. hf cache prune deletes revisions that no branch or tag points to any more, partial files from interrupted downloads, and shared files no cached repository uses:
hf cache prune
With older versions, huggingface-cli delete-cache opens a list where you tick what to remove.
The CLI is safer than Finder here. Recent versions store some files once, in a shared blobs folder at the top of hub, for every repository that uses them, and only hf cache rm and hf cache prune know when those can go. The xet folder is the exception: Hugging Face’s docs say you can remove it entirely to reclaim its space.
5. What happens afterwards
- The next
from_pretrained()orload_dataset()that needs a removed repository downloads it again. That needs the internet and, for gated models, your token. - Code running offline, with
HF_HUB_OFFLINE=1set, only reads the cache, so it raises an error for anything you removed. - Repositories you kept, and their revisions, are untouched, and your token stays where it was.
Questions
Can I delete the whole ~/.cache/huggingface folder?
You can, with no Python process using it, but you also delete your login token. Removing hub and datasets and keeping the rest gives back almost the same space without signing in again.
Why is the cache bigger than the models I use?
Every revision you’ve downloaded stays, and so do models from old experiments. hf cache ls --revisions shows them, and hf cache prune removes revisions nothing points to.
Does deleting the cache remove models I fine-tuned?
Not if you saved them to your own folder with save_pretrained(). Only files downloaded from the Hub live in the cache.
Is this the cache LM Studio and Ollama use?
No. LM Studio and Ollama keep their own copies of models, outside ~/.cache/huggingface.
How SiftMac handles it. Its AI files category lists each model folder in ~/.cache/huggingface/hub (models--org--name) with its size and when it last changed. Models are You decide and never ticked for you, and one you choose is removed straight away rather than kept on the Undo Shelf; datasets and the rest of the cache are left alone.
Download SiftMac · scanning is free, and so is your first 2 GB of cleaning · See how it works