Tidlark

Guide

How to remove Ollama, LM Studio and Hugging Face models on a Mac

By Maulik Dave · Updated

Local models are big: a small one is a few gigabytes, a large one can be 40 GB or more. Each tool stores them in its own place, often hidden. Here’s where they are on a Mac and how to remove them with each tool’s own commands.

Short answer: on a Mac, Ollama stores models in ~/.ollama/models (a manifests folder and a blobs folder), unless the OLLAMA_MODELS environment variable points elsewhere. On Linux, the standard installer uses /usr/share/ollama/.ollama/models. Remove a model with ollama rm, not by deleting blobs (Ollama FAQ).

Use the tool’s own command where there is one. Ollama and Hugging Face store model data as shared “blob” files: one blob can belong to several models. Deleting blob files by hand can break models you meant to keep, and the space may not even be freed.

Where are Ollama models stored on a Mac?

Where: on macOS, Ollama stores models in ~/.ollama/models. Inside, manifests describes each model and blobs holds the data. If you set the OLLAMA_MODELS environment variable, models are in that folder instead. The Ollama FAQ says environment variables for the Mac app are set with launchctl setenv, followed by a restart of Ollama:

launchctl setenv OLLAMA_MODELS /Volumes/External/ollama-models

On Linux the default is /usr/share/ollama/.ollama/models, and on Windows C:\Users\%username%\.ollama\models.

List your models with their sizes:

ollama ls

ollama list does the same. Remove one by the name shown in that list, including its tag (see the Ollama CLI reference):

ollama rm llama3.2:3b

Ollama deletes the model’s manifest and any blobs that no other installed model uses. Removing a model can’t be undone, but you can always download it again with ollama pull.

If a model is loaded in memory, ollama ps shows it and ollama stop <model> unloads it. Stopping frees memory, not disk space.

To check how much space Ollama uses in total:

du -sh ~/.ollama/models

To remove Ollama itself, quit it from the menu bar and move the app to the Trash. Its models stay in ~/.ollama until you remove that folder too; run ollama rm for each model first if you might reinstall, or move the whole ~/.ollama folder to the Trash in Finder if you’re done with Ollama.

LM Studio

Where: LM Studio’s documentation gives the models directory as ~/.lmstudio/models, organised as publisher/model/model-file.gguf, mirroring the Hugging Face repository the model came from. You can move the directory from the My Models tab in the app, so check there if yours is elsewhere.

List your models from the app’s My Models tab, or in Terminal with LM Studio’s CLI (lms ls docs):

lms ls

Remove one by deleting it from My Models in the app. Or quit LM Studio, open the models folder in Finder (⇧⌘G, then ~/.lmstudio/models) and move the model’s folder to the Trash. The folders hold plain model files, so moving one to the Trash removes just that model, and you can put it back until you empty the Trash.

Hugging Face cache

Where: libraries that download from the Hugging Face Hub (transformers, diffusers, MLX tools and others) share one cache. Its default location is ~/.cache/huggingface/hub (Hugging Face cache guide). The HF_HOME or HF_HUB_CACHE environment variable moves it. Each repository gets a folder like models--org--name with blobs (the data) and snapshots (links to blobs, one folder per revision).

The Hugging Face hf command-line tool, part of the huggingface_hub Python package, manages the cache. List what’s cached, largest first:

hf cache ls --sort size:desc

Preview a removal, then run it without --dry-run; it asks you to confirm:

hf cache rm model/org/name --dry-run
hf cache rm model/org/name

Use the IDs exactly as hf cache ls shows them (for example model/bert-base-cased). A blob is deleted only when no other cached file still uses it. To remove a single file, such as one GGUF quantization, pass its hf:// path instead:

hf cache rm hf://models/org/name/model-Q4_K_M.gguf --dry-run

To clear revisions no branch or tag points to any more, leftovers of interrupted downloads and shared files no cached repository uses:

hf cache prune --dry-run

Hugging Face recommends an up-to-date huggingface_hub for hf cache rm and hf cache prune (pip install -U huggingface_hub). Deleting a cached model is permanent; it downloads again the next time code asks for it, which for gated models needs your access token.

If you upload to the Hub, ~/.cache/huggingface/xet also holds data used to speed up and resume uploads. Hugging Face’s cache guide says you can remove that xet folder entirely if you need the space; downloaded models are not stored there.

Other tools

GPT4All, Jan, llama.cpp, ComfyUI, Draw Things and others each keep models in their own folders, usually under ~/Library/Application Support or a folder you chose. Check each app’s settings for its model location and delete models from inside the app where you can.

If you’re cleaning up for space in general, see how to free up disk space on a Mac. Developers will also want how to delete Xcode DerivedData.

Where Tidlark helps

Tidlark’s AI Models & Files section finds models from Ollama, Hugging Face, LM Studio, GPT4All, Jan, llama.cpp, MLX, ComfyUI, AUTOMATIC1111, Draw Things and DiffusionBee, plus any folder you add. For each model it shows:

  • the tool, size on disk and last-modified date;
  • format and quantization, read from the file’s own header (GGUF, safetensors) rather than guessed from the name; unknown values are shown as unknown;
  • whether it shares files with another model. Shared files are counted once and are never removed while another model still uses them. Selecting a shared model shows a warning first;
  • older versions of the same model, with a button that selects them and keeps the newest.

A separate Caches & Leftovers tab lists logs, interrupted downloads and blobs no installed model uses. A download that may still be in progress can’t be removed. Nothing is selected for you, and Tidlark re-checks every file just before removal.

See pricing or download Tidlark. Related: uninstall apps completely.