How to clear the Hugging Face cache

Every from_pretrained() and load_dataset() call caches files on disk, and large models make that folder grow fast. The hf CLI can list and delete per model, which is gentler than wiping the folder.

Where it lives

• macOS / Linux: ~/.cache/huggingface/hub (models), ~/.cache/huggingface/datasets (processed datasets)
• Windows: C:\Users\<you>\.cache\huggingface\hub

HF_HOME or HF_HUB_CACHE override these.

List what is cached

pip install -U huggingface_hub
hf cache ls
hf cache ls --revisions

huggingface-cli scan-cache

Delete models or old revisions

hf cache rm model/bert-base-uncased --dry-run
hf cache rm model/bert-base-uncased
hf cache prune --dry-run   # detached revisions + unfinished downloads
hf cache prune

huggingface-cli delete-cache   # interactive picker

Run with --dry-run first to see exactly what goes.

Windows note

Without Developer Mode (or admin rights) the cache cannot use symlinks, so files are duplicated and the cache uses more space than on macOS or Linux.

Measure with devdisk

devdisk is a free, read-only scanner. It lists ~/.cache/huggingface under Review first (AI models), since re-downloading big models is slow. It never deletes anything; it only shows sizes and writes a cleanup script you read and run yourself.

curl -fsSLO https://disk.sadana.app/devdisk.py && python3 devdisk.py

iwr https://disk.sadana.app/devdisk.py -OutFile devdisk.py; py devdisk.py

Run free scannerSample report

FAQ

More cleanup guides

• Clear Xcode DerivedData
• Delete node_modules
• Clear Gradle cache
• Delete Ollama models
• Free Docker disk space
• Clear Homebrew cache
• Prune pnpm store
• Clear Yarn cache
• Clean Rust target/
• Clear pip & uv cache
• Clear .next / .turbo
• Delete Android emulators