Where AI Models Take Disk Space on a Mac

If you run AI models on your Mac, most of the space they take is in three places: Ollama keeps its models in ~/.ollama/models, LM Studio in the models folder shown on its My Models tab, and Python tools that download from Hugging Face in ~/.cache/huggingface. A single model is often several gigabytes, and they are easy to forget once you have tried them. AI apps such as Claude, ChatGPT and Cursor add caches of their own. Below is where each one lives, how to see its size, and how to remove models with each tool's own command, which is safer than deleting files by hand.

First, see how much space they use

1. In Finder choose Go > Go to Folder and type ~/.ollama. If the folder opens, Ollama has stored models.
2. Select the models folder and choose File > Get Info to see its size. Do the same for ~/.cache/huggingface and for your LM Studio models folder.
3. These folders start with a dot, so they are hidden in Finder. Go to Folder opens them anyway. You can also press Cmd + Shift + . in a Finder window to show hidden files.

System Settings > General > Storage does not list these folders by name, which is why they are easy to miss there.

Ollama

On a Mac, Ollama stores models in ~/.ollama/models. If you set the OLLAMA_MODELS environment variable, they are wherever that points instead.

1. Open Terminal and run ollama ls to list the models you have, with their sizes.
2. Remove one you no longer use with ollama rm followed by its name, for example ollama rm llama3.2.

Use the command rather than deleting files in the blobs folder. Ollama stores models as layers, and two models can share a layer, so a file you delete by hand may still be needed by a model you are keeping. To keep models on an external drive instead, set OLLAMA_MODELS to a folder there; Ollama's FAQ explains how to set it for the Mac app.

LM Studio

1. Open LM Studio and go to the My Models tab. It lists the models on disk and shows the folder they are kept in. On most Macs that is ~/.lmstudio/models, but the folder shown there is the one to trust.
2. To remove a model, quit LM Studio, open that folder in Finder and move the model's folder to the Trash. Models are kept in one folder per publisher, with one folder per model inside.
3. To move all models to another drive, change the models folder on the same tab.

If you use the lms command line tool, lms ls lists the models in that folder. LM Studio also keeps caches in ~/Library/Application Support/LM Studio, which are smaller.

Hugging Face

Libraries that download from Hugging Face, such as transformers and diffusers, keep a shared cache in ~/.cache/huggingface/hub, unless HF_HOME or HF_HUB_CACHE points elsewhere. Each model is a folder named like models--org--name. Newer versions also keep a xet folder beside it.

1. Run hf cache ls to list what is cached and how big each item is.
2. Remove a model with hf cache rm. Add --dry-run first to see what would go.
3. hf cache prune removes old revisions no longer in use and unfinished downloads.

The hf command comes with the huggingface_hub Python package. Older guides use huggingface-cli; the current Hugging Face documentation shows hf.

AI apps: Claude, ChatGPT, Cursor

Desktop AI apps keep caches like any other app, and the coding tools keep history too:

• Claude (the desktop app) keeps caches in ~/Library/Application Support/Claude, in folders named Cache, Code Cache and GPUCache.
• Cursor keeps caches in ~/Library/Application Support/Cursor (Cache, CachedData, Code Cache, GPUCache) and extensions in ~/.cursor/extensions.
• ChatGPT keeps its cache in ~/Library/Caches/com.openai.chat, or inside ~/Library/Containers/com.openai.chat.
• Claude Code and Codex (the command line tools) keep session history in ~/.claude/projects and ~/.codex/sessions, which grow with use.

Quit the app before you clear its cache, and remove only the cache folders. The app rebuilds them. Never delete ~/.claude or ~/.codex as a whole: they hold your settings and sign-in. Session history is your data, not a cache, so remove it only if you no longer need old conversations.

What MacPrune shows for AI tools

We make MacPrune, and its AI page exists for exactly this. It looks in the places above for Ollama, LM Studio, Hugging Face, Claude, Claude Code, ChatGPT and Codex, and Cursor, and lists each model, cache and data folder with its size. Models downloaded more than 90 days ago are marked as old. For Ollama it lists one entry per model and never offers a layer that another model still uses; layers no model refers to, such as those left by an interrupted download, are listed separately. It never lists the folders that hold settings or sign-in.

The same page shows AI processes using CPU and memory, GPU use, and local AI servers listening on ports, such as Ollama's and LM Studio's, with a way to quit a process or stop a server. Anything you choose to clean moves to the Trash after a confirmation; if the tool is still running, MacPrune offers to quit it first. The free version clears AI caches. Removing models and app data needs Pro, a one-time purchase; see pricing.

Models are often the largest thing on a developer's Mac, next to Xcode. For that, see Xcode taking up storage, and for every other place space hides, how to clear System Data and how to free up Mac storage. You can download MacPrune to see the AI page with your own tools.