How to delete local AI models on a Mac: Ollama, LM Studio and Hugging Face
Local AI models are some of the biggest files on a Mac, and they're easy to forget. Here's where each tool keeps them and how to remove them properly. Updated 14 November 2026.
Why models take so much space
A local model is the full set of weights an app runs on your Mac, stored as one or more large files. A single model often takes several gigabytes, and it's common to try a few sizes or versions of the same model and keep them all. Unlike a cache, a model isn't rebuilt from anything on your Mac. If you remove one and want it back, you download it again.
Each tool keeps its own copy in its own folder. If you've used Ollama, LM Studio and a Python library from Hugging Face, you may have the same model stored three times.
First, see where the space is
These Terminal commands show how big each tool's default folder is. They only read sizes and don't change anything:
du -sh ~/.ollama/models
du -sh ~/.lmstudio/models
du -sh ~/.cache/huggingface
"No such file or directory" means that tool hasn't stored anything in its default place. Folders whose names start with a dot are hidden in Finder. Press Command-Shift-. (period) in a Finder window to show them.
Ollama
On a Mac, Ollama keeps models in ~/.ollama/models, unless you've set the OLLAMA_MODELS environment variable to put them somewhere else (Ollama FAQ). Remove them with Ollama's own commands (Ollama CLI reference):
- List what you have:
ollama ls(orollama list) shows each model's name, size and when it last changed. - Make sure it isn't running:
ollama pslists loaded models, andollama stop <model>stops one. - Remove it:
ollama rm <model>, using the name exactly asollama lsshows it, tag included. - Check: run
ollama lsagain, and theducommand above.
Don't delete files inside ~/.ollama/models by hand. The folder holds manifests, which describe each model, and blobs, which hold the large files. Models built on the same base can share blobs, so a file you delete in Finder may belong to a model you still use. ollama rm works out which blobs nothing else needs. That sharing is also why removing one model can free less than the size ollama ls showed for it.
To remove Ollama completely, not just its models, Ollama's macOS page lists every file and folder to remove, including ~/.ollama (Ollama on macOS).
LM Studio
LM Studio keeps downloaded models in its models folder, which is ~/.lmstudio/models by default in current versions. You can move it, so the reliable way to find yours is the My Models tab in the app, where the models directory is set (LM Studio docs).
- Unload it first if it's in use. In Terminal,
lms pslists the models loaded in memory andlms unload --allunloads them. Unloading frees memory, not disk space. - Remove it in the app. Open My Models, find the model by name and size, and use its delete option. LM Studio removes the files and updates its list.
- Check:
lms lslists the models still on disk (lms reference).
The lms command comes with LM Studio. If Terminal can't find it, do everything in the app. Prefer the app to Finder either way: it knows which files belong to which model.
Hugging Face cache
Python libraries that download models from Hugging Face share one cache. By default it's ~/.cache/huggingface/hub, or wherever the HF_HOME or HF_HUB_CACHE environment variable points. Hugging Face's own documentation notes that cached files are never removed on their own: when you download a newer revision of a model, the old files are kept too (Manage the Hugging Face cache).
The hf command, which comes with the huggingface_hub package, handles it:
hf cache ls --sort size:desclists each cached model and dataset with its size and when you last used it, biggest first.hf cache rm model/<org>/<name> --dry-runshows what would be removed and how much space it frees. Run it again without--dry-runto remove it. It asks before it starts.hf cache pruneremoves old revisions that no branch or tag points to any more, plus downloads that were interrupted. It also asks first, and takes--dry-runtoo.
Older versions of the package use huggingface-cli scan-cache and huggingface-cli delete-cache instead. pip install -U huggingface_hub updates it.
Finder can mislead you here. Each model's snapshots folder holds links rather than files, so it looks tiny, while the real data sits in blobs. The hf commands count it properly.
Other apps that download models
Other apps that run models on your Mac keep them in their own places, usually in ~/Library/Application Support or in a hidden folder in your home folder. Look for a models or storage section in the app's settings and remove models there, so the app knows they're gone. If an app has no way to remove a model, check its documentation before deleting its folders.
Which models to keep
- Keep what you run. If you have the same model in several sizes or quantizations, the one you actually use is usually enough.
- Remove copies across tools. If Ollama and LM Studio both hold the same model, keep it in the tool you use most.
- Think about your connection. Getting a model back means downloading all of it again, which can take a long time on a slow or metered connection.
- Back up your own work first. Models you fine-tuned, merged or converted yourself exist only on your Mac. No command brings those back.
What CheckMac does here
CheckMac shows model folders on the disk map at their real size, including hidden folders like ~/.ollama and ~/.cache (hidden files are shown by default), so one scan of your home folder tells you how much each tool is holding. It never suggests models for cleanup: none of its rules covers ~/.ollama, LM Studio's models folder or ~/.cache/huggingface, because a model isn't rebuilt for free. If a tool keeps a cache in ~/Library/Caches, such as an ollama or huggingface folder there, Suggested Cleanup can list it as an app cache, but it marks it as costly to rebuild and never selects it for you. Remove models with each tool's own command, as above. For the rest of a developer Mac, see the developer checklist, and for Docker's disk image, Docker.raw explained. If you're not sure AI models are the problem at all, start with how to see what's taking up space on your Mac.


See it on your own Mac. CheckMac maps any folder or drive as a sunburst or treemap, and Suggested Cleanup lists the caches and build output a tool can rebuild, with what each one costs you. Scanning is free, and Suggested Cleanup is free for its first three runs.
More guides
- What is System Data on Mac, and how do you reduce it?
- How to delete Xcode DerivedData, simulators and device support files safely
- Is it safe to delete node_modules, caches and build folders on a Mac?
- Looking for a CleanMyMac alternative?
- Mac storage almost full on a 256 GB Mac: what to do
- How to see what's taking up space on your Mac
- Free up disk space on a Mac for developers: a checklist
- Docker taking up space on Mac: Docker.raw explained
- How to delete old iOS simulators and simulator runtimes
- Purgeable space on a Mac: what it is and how to clear it