Ollama and Hugging Face are eating your disk

By the maker of WTMemory · Updated Sept 2026

You ran a 7B model once to see what the fuss was about. Six months later your Mac is out of space and Finder blames "System Data". Here is where local AI actually lives on macOS, how much it costs you, and how to reclaim it without breaking your setup.

Where AI models hide (with typical sizes)

Tool	Path	Typical size
Ollama	~/.ollama/models	4-60 GB
Hugging Face cache	~/.cache/huggingface	5-100 GB
LM Studio	~/.cache/lm-studio/models	10-80 GB
ComfyUI / Diffusion	~/ComfyUI/models and checkpoints	20-200 GB
Cursor / editor indexes	~/Library/Application Support/Cursor/User/workspaceStorage	1-20 GB

Find your own numbers in one command:

du -sh ~/.ollama ~/.cache/huggingface ~/.cache/lm-studio 2>/dev/null

Is it safe to delete them?

Yes - with one caveat. Model weights are downloads, not data. Deleting them means the tool re-downloads a model the next time you reference it, which costs bandwidth and a few minutes, nothing else. Your prompts, chats and settings live elsewhere and are untouched.

The caveat: if you are on metered or slow internet, deleting a 40 GB model you use weekly is a bad trade. Keep what you actually run; delete the graveyard of experiments.

Clean them selectively, not wholesale

# See what each Ollama model weighs
ollama list

# Remove just one
ollama rm llama3.3:70b

# Prune the Hugging Face cache to only what is referenced recently
pip install huggingface_hub
huggingface-cli delete-cache

For LM Studio and ComfyUI, models are individual files - delete them from the app's model browser or Finder, and delete the folder rather than trying to surgically trim it.

The one thing not to touch

Do not delete the whole ~/.ollama or ~/.claude directory to "save space." The first contains your model registry and config; the second contains agent settings, sessions and API credentials. Kill the models, keep the tool.