DiskBroom › What it cleans › Hugging Face cache taking up space? Here's the fix

The Hugging Face cache grows until you look at it.

~/.cache/huggingface

What it is

Every model, dataset, and tokenizer that transformers, diffusers, or huggingface_hub downloads is cached here — every version of each, kept indefinitely so the next run is instant.

How big it gets

A single modern model checkout runs 2–15 GB; anyone experimenting with a few models commonly reaches 20–100 GB without noticing.

What happens if you delete it

Anything removed re-downloads automatically the next time a script asks for it. The cost is bandwidth and wait, not data — your own fine-tunes and outputs live wherever you saved them, not in the cache.

How to remove it manually — free, no app needed

1. See the total: du -sh ~/.cache/huggingface
2. Best tool: pip install huggingface_hub[cli], then huggingface-cli delete-cache — an interactive picker that shows each cached repo and its size.
3. Manual route: delete individual model folders under ~/.cache/huggingface/hub (each is named after its repo).
4. Nuclear route: delete the whole ~/.cache/huggingface folder; everything re-downloads on demand.

Risk rating

Worth a look. Safe in the sense that everything re-downloads — but multi-GB re-downloads deserve a glance first.

Something missing, or wrong in this guide? Request a feature · Report a problem

More cleaning guides

• How to delete Xcode DerivedData safely
• What "System Data" on Mac actually is
• Where Ollama stores models (and how to remove them)
• How to safely clear npm, pnpm, and Yarn caches
• Docker Desktop using huge disk space on Mac
• Adobe Media Cache: why it's huge and how to clear it
• Clearing the Homebrew cache (brew cleanup, explained)

← Everything DiskBroom can clean · FAQ