Why does my Mac still show low storage after I deleted an AI app?

The app bundle is almost never the storage problem. Model files live in hidden folders like ~/.ollama/models/ that macOS Storage Management doesn't show under Applications. Use the du -sh command from the checklist above to find them.

Where are GGUF model files stored?

Ollama: ~/.ollama/models/blobs/ (files named by content hash, not model name). LM Studio: ~/.cache/lm-studio/models/ (organized by publisher and model name). GPT4All: ~/Library/Application Support/nomic.ai/GPT4All/.

How much RAM does local AI need on a Mac?

On Apple Silicon, RAM and VRAM are the same pool, so the number matters a lot: 7-8B models need 16 GB recommended (8 GB works but causes swapping), 13B models need 24 GB, and 70B models need 48 GB or more.

Can I move models between Ollama and LM Studio?

Both use GGUF files, but Ollama stores blobs by content hash while LM Studio uses human-readable folder names. You can't just copy Ollama blobs into LM Studio's folder and have them work. The practical approach: download the model fresh in the new tool, then delete it from the old one.

Why is LLM performance so poor on my Intel Mac?

Most local LLM shells run CPU-only inference on Intel Macs, since discrete GPU VRAM is usually too small. For 7B+ models, that means slow generation speeds. It's one of the main reasons Intel Mac users end up removing AI tools entirely.

June 1, 2026

AI on Mac: A complete checklist for removing AI tools leftovers

Written by

A Mac specialist with a QA engineering background, focused on troubleshooting and how-to guides.

Sergio Tereshchenko

Approved by

Reviewed and approved by an Apple Certified Support Professional and Mac developer with hands-on experience supporting Nektony users.

Vladyslav Zubkov

Running LLMs locally sounds like a miracle: no connection required, your data is safe. But that miracle disappears when you switch tools, run out of RAM or space, you’re not sure where all those files are. The app is hundreds MB. The models hiding behind it are 40+GB. macOS won’t show you either number in the right place.

This piece of content is about the nuances of having, handling, and uninstalling every major AI app (local LLM shells, cloud desktop clients, browser tools) and links to full removal guides for each one. Because uninstalling an app is simple, but not an AI one.

This Article Contains

Types of AI tools on Mac
- Why AI apps take so much space
- Why uninstalling AI tools is different
Ollama
LM Studio
Claude Desktop and Claude Code
ChatGPT Desktop
Other cloud AI clients
Other local AI tools
AI tools in browsers
Final cleanup checklist
FAQs

Types of AI tools on Mac

Before you delete anything, it helps to know what you’re actually dealing with. AI tools on Mac split into three categories with very different footprints.

Type	Local LLM shells	Cloud AI clients	AI coding agents/CLIs
Tool	Ollama, LM Studio, Pinokio, Anything LLM, Jan, GPT4All	Claude, ChatGPT, Gemini, Perplexity	Claude Code, OpenAI Codex, Aider, Gemini
Where it runs	Fully on your Mac	Mostly on remote servers	Local terminal + cloud
What on disk	App + model files (GBs to tens of GBs)	App + cache, logs, history	Configs, repos, logs, auth tokens, caches
Privacy	Prompts never leave your environment	Prompts sent to provider	Local files may be sent to provider or indexed

Local LLM shells download model weight files to your Mac and run them locally using an inference engine like llama.cpp or MLX. The app itself is small; the models are what fill your drive. One model in Q4 quantization runs 4-40 GB, depending on size. Test five or six, and you can easily accumulate 50-100 GB.
Cloud AI desktop clients are thin interfaces that send your prompts to remote servers. They store UI caches, conversation history, auth tokens, and configs locally, but no model files. Removing them is closer to a normal Mac app uninstall, though there are still hidden folders that a drag-to-Trash won’t catch.
AI coding agents/CLI tools are something in the middle. These tools run locally in Terminal but rely on cloud models for inference. They integrate deeply with your filesystem, repos, shell history, MCP servers, and dev environments.

Deleting the binary often leaves behind configs, caches, auth tokens, project sessions, npm/Homebrew package data, and sometimes multiple parallel installations hidden in different Node.js environments.
Hybrid cases: Notion AI and Raycast AI are AI features built into larger apps. There’s no way to remove just the AI component, you can disable it in settings or uninstall the whole app.

Why AI apps take so much space

The app file is almost never the storage problem. Here’s what typical model files actually weigh:

Model	Format	Size
Mistral 7B	Q4 GGUF	~4.1 GB
Llama 3.1 8B	Q4 GGUF	~4.7 GB
Mixtral 8x7B	Q4 GGUF	~26 GB
Llama 3.1 70B	Q4 GGUF	~40 GB

The models are buried in hidden dot-folders that macOS System Settings → General → Storage does not count under Applications. This is a documented macOS limitation: Storage Management only indexes certain directories, and

~/.ollama/models/

Copy

isn’t one of them. So the storage readout looks fine while tens of gigabytes sit invisible.

Run this in Terminal to see what AI tools are actually using:

du -sh ~/.ollama ~/.cache/lm-studio ~/jan ~/Library/Application\ Support/nomic.ai 2>/dev/null

Copy

AI model file formats

Three formats come up when you look at what’s on disk:

GGUF: used by Ollama, LM Studio, GPT4All. The dominant format for quantized local models.
safetensors: the standard HuggingFace format, used by some tools and fine-tuning workflows.
MLX: Apple Silicon–optimized format used by mlx-lm and some LM Studio variants.

One thing to watch for: the same model can exist in multiple formats at once. If you downloaded Llama 3.1 8B in GGUF for Ollama and in MLX for another tool, you’re storing what is functionally the same model twice, around 10-15 GB total.

RAM and Apple silicon

On Apple Silicon, RAM and VRAM share the same unified memory pool. Running Llama 3.1 8B on an M1 with 8 GB leaves almost nothing for macOS itself; the system compensates by swapping to the SSD, which slows everything down and adds unnecessary write cycles. In practice:

8 GB: workable for small models with nothing else open, but tight
16 GB: comfortable for 7-8B models
24-48 GB: needed for 13B and above

On Intel Macs, most shells run models through CPU inference only, which is slow enough for 7B+ models that many Intel Mac users remove AI tools for exactly this reason. The storage paths are identical, so this guide applies to both architectures.

Why uninstalling AI tools is different

Aspect	Local LLM shells	Cloud AI clients	CLIs
Removal complexity	High: hidden model folders, LaunchAgents	Moderate: hidden support folders	High: configs, package manager leftovers
Common uninstall issue	Models stay after deleting the app	Cache and support files remain in ~/Library	CLI removed but configs/history remain
Typical storage use	4-80+GB models stored locally	100MB-2GB mostly caches, logs, chat history	50MB-5+GB, depends on repos, sessions, VM

With a regular Mac app, removal is pretty clean. Delete the app executable file → clear out main folders like Application Support, Containers, Caches → and you’re done. AI tools don’t follow that pattern. As stated above, most of their files are invisible.

Item	Regular Mac app	AI tool
App bundle	`/Applications/App.app` Copy (~50–500 MB)	`/Applications/Ollama.app` Copy (~500 MB)
App data	`~/Library/Application Support/App/` Copy	`~/.ollama/models/` Copy (4-80+ GB)
Cache	`~/Library/Caches/App/` Copy	`~/Library/Caches/` Copy + `~/.cache/` Copy
Background service	Usually none	LaunchAgent plist (may restart itself)
System binary	Usually none	`/usr/local/bin/ollama` Copy
Visible in Storage?	Yes	Partially, model folders often hidden

Drag an AI tool like Ollama to the Trash and you’ll reclaim about 500 MB, leaving Library hidden folders untouched.

The LaunchAgent trap.

Ollama and several other AI tools register a background service that starts at login. If you delete the app without first unloading and removing that LaunchAgent, the service tries to restart a binary that no longer exists, and generates errors in system logs at every login. Always unload the agent before removing the plist.

Check which AI-related LaunchAgents are currently registered:

ls ~/Library/LaunchAgents/ | grep -E ’ollama|lmstudio|jan|gpt4all’

Copy

Ollama

Ollama is the most widely used shell for running local LLMs on Mac. You can install it from the official website or via Homebrew (

brew install ollama

Copy

It runs as a background service, exposes a local API endpoint, and handles model downloads through simple

ollama pull

Copy

commands. What it doesn’t do is clean up after itself when you delete the app.

What Ollama stores

Path	Contents
`/Applications/Ollama.app` Copy	App bundle
`~/.ollama/models/blobs/` Copy	Model weight files (not named by model)
`~/.ollama/models/manifests/` Copy	Index of downloaded models
`~/.ollama/history` Copy	CLI command history
`/usr/local/bin/ollama` Copy	CLI binary
`~/Library/LaunchAgents/com.ollama.ollama.plist` Copy	Auto-start service

The blobs folder is where all the space is. Filenames are content hashes, not model names, you can’t tell

mistral-7b

Copy

from

llama-3.1-8b

Copy

by looking at filenames alone. Use

ollama list

Copy

to see what’s installed while the app is running.

Note:

brew uninstall ollama

Copy

only removes the binary. It does not delete

~/.ollama/models/

Copy

or touch the LaunchAgent. If you installed via Homebrew, you still need to clean up the rest manually.

Full removal steps: How to uninstall Ollama and models on Mac

LM Studio

LM Studio is a GUI app for downloading and running GGUF models locally. It also runs a local server with an OpenAI-compatible API. Unlike Ollama’s content blob storage, LM Studio organizes models by publisher and model, which makes them easier to browse.

What LM Studio stores

Path	Contents
`/Applications/LM Studio.app` Copy	App bundle
`~/.cache/lm-studio/models/` Copy	Downloaded model files (organized by publisher/name)
`~/.cache/lm-studio/` Copy	App cache, inference logs
`~/Library/Application Support/LM-Studio/` Copy	App settings and data

GGUF files are portable. If you’re moving from LM Studio to Ollama (or elsewhere), you can point the new tool at the existing model files instead of re-downloading. That said, the folder structure each tool expects is different, so it’s not plug-and-play, but it’s worth knowing before you delete everything and start over.

Removing the LM Studio app without clearing

~/.cache/lm-studio/models/

Copy

and other support files and folders leaves the bulk of the data behind.

Full removal steps: How to uninstall LM Studio on Mac completely

Claude Desktop and Claude Code

Claude Desktop (

com.anthropic.claudefordesktop

Copy

) is Anthropic’s native macOS app. It doesn’t store model files locally, but it has one storage detail that sets it apart from other cloud clients: MCP (Model Context Protocol) server configurations are saved in a local JSON file that can contain API keys, file paths, and sensitive tokens.

Claude Code (

com.anthropic.claude-code

Copy

) is another developer CLI tool that uses Claude Desktop as its host. The two share

~/Library/Application Support/Claude/

Copy

, so their files overlap on disk. Removing one can affect the other.

That’s why you have to be careful with how you remove them. Deleting both is one scenario, and removing one of them is another scenario, where it’s important to check what still relies on the shared folder so nothing breaks.

What Claude Desktop stores

Path	Contents
`~/Library/Application Support/Claude/` Copy	App data (shared with Claude Code)
`~/Library/Application Support/Claude/vm_bundles/` Copy	VM images for Claude Code (~8 GB)
`~/Library/Caches/com.anthropic.claudefordesktop/` Copy	UI cache
`~/Library/Logs/Claude/` Copy	App logs
`~/.config/claude/claude_desktop_config.json` Copy	MCP server config — may contain API keys
macOS Keychain	Auth token (entry named “Claude”)

What Claude Code stores

Path	Contents
`~/Library/Application Support/Claude/claude-code/` Copy	Executable (inside shared folder)
`~/.claude/` Copy and `~/.claude.json` Copy	CLI settings and memory
`~/Library/Caches/claude-cli-nodejs/` Copy	Node.js runtime cache

The

~/Library/Application Support/Claude/

Copy

folder can grow significantly. The

vm_bundles/

Copy

subfolder averages around 8 GB; the rest of Claude Desktop’s data typically runs 1.5-2 GB.

Note:

~/.config/claude/claude_desktop_config.json

Copy

may contain API keys and tokens for connected MCP servers in plain text. When selling or transferring a Mac, delete this file explicitly, removing the app alone won’t touch it.

ChatGPT Desktop

ChatGPT Desktop is OpenAI’s native macOS app. No local model files kept. The storage is conversation history, UI cache, and app data.

The storage location depends on how you installed it:

Install source	App data path
Direct download	`~/Library/Application Support/ChatGPT/` Copy
Mac App Store	`~/Library/Containers/com.openai.chat/` Copy

Both versions also write to

~/Library/Caches/com.openai.chat/

Copy

. The App Store version adds data in

~/Library/Group Containers/

Copy

under OpenAI’s app group ID.

Removing ChatGPT from the App Store via Launchpad (long-press → X) removes the app but leaves

~/Library/Containers/com.openai.chat/

Copy

untouched. This is standard macOS behavior for sandboxed App Store apps, container data is only cleaned up if you use the uninstaller like App Cleaner & Uninstaller or delete the folder manually.

Other cloud AI clients

Perplexity

Perplexity stores conversation history and preferences in

~/Library/Application Support/Perplexity/

Copy

and cache in

~/Library/Caches/

Copy

. There are no model files. Standard uninstall plus folders in the Library handles it completely.

Gemini

Gemini on Mac runs as a PWA (Progressive Web App) installed through Chrome (Settings → Cast, Save, and Share → Install Page as App). Its data is inside Chrome’s profile at

~/Library/Application Support/Google/Chrome/Default/

Copy

To remove: open Chrome → go to

chrome://apps

Copy

→ right-click Gemini → Uninstall → check the box next to Remove this app’s data from Chrome.

Chrome pop-up to uninstall Gemini with its data

Built-in AI features

Raycast AI is a feature inside the Raycast launcher, not a separate app. Disable it under Raycast Settings → AI, or uninstall Raycast entirely from

/Applications/

Copy

Notion AI is built into the Notion desktop app. Disable it in Notion Settings → AI, or uninstall Notion if you want a full removal.

Other local AI tools

Pinokio

Pinokio stores all installed AI apps, models, and dependencies in

~/pinokio/

Copy

, which is the main folder that can easily reach 80+ GB.

Additional leftovers are:

```
~/.agents/skills/pinokio/
```
Copy
```
~/.agents/skills/gepeto/
```
Copy
```
~/Library/Application Support/Pinokio/
```
Copy
```
~/Library/Caches/computer.pinokio
```
Copy

The

~/.agents/

Copy

folder is shared with other AI tools. Delete only the

pinokio/

Copy

and

gepeto/

Copy

subfolders, not the entire directory.

AnythingLLM

AnythingLLM saves all data (docs, vector embeddings, chat history, and settings) in a single folder:

~/Library/Application Support/anythingllm-desktop/

Copy

. Models themselves are typically pulled from Ollama or LM Studio, so no separate model files. Standard uninstall plus deleting that one folder handles it completely.

The official guide includes a dedicated Uninstall page.

GPT4All

GPT4All keeps downloaded models in

~/Library/Application Support/nomic.ai/GPT4All/

Copy

, where models are several GB each. Chat history and settings are in the same folder. Standard uninstall plus that folder is a complete removal. It’s one of the simplest cleanups in this category.

Jan

Jan stores models in

~/jan/models/

Copy

and all app data (threads, settings, extensions) in

~/jan/

Copy

. This is Jan’s custom home directory outside the standard Library paths, so it’s easy to miss. Standard uninstall leaves the entire

~/jan/

Copy

folder intact. No Library folders to worry about, just delete

~/jan/

Copy

and you’re done.

AI tools in browsers

Using ChatGPT, Claude, or Gemini in a browser without any desktop app still leaves data on disk. Here’s what browsers store:

Cookies – session tokens for
```
chatgpt.com
```
Copy
,
```
claude.ai
```
Copy
,
```
gemini.google.com
```
Copy
localStorage – UI preferences, conversation drafts, cached history
IndexedDB – structured conversation data in some implementations
Service Workers/Cache API – offline components, especially for Gemini

Clearing browser data by browser

Safari

Go to Safari → Settings → Privacy → Manage Website Data → search for the site → Remove.

For a full wipe: History → Clear History removes cookies, history, and cache together. Safari cache files are in

~/Library/Caches/com.apple.Safari/

Copy

Chrome

Open Settings → Privacy and security → Delete browsing data → check Cookies and Cached images. For one site: click the lock icon in the address bar → Site settings → Clear data. Chrome’s profile data is in

~/Library/Application Support/Google/Chrome/Default/

Copy

Firefox

Go to Settings → Privacy & Security → Cookies and Site Data → Manage Data → find the domain → Remove Selected.

Service workers

Gemini registers a Service Worker that persists after you close the tab. To remove it in Chrome: open DevTools (Command+Option+I) → Application → Service Workers → Unregister.

AI browser extensions

Disabling an extension in Safari Settings → Extensions does not remove its stored data, you need to click Uninstall. Some older extensions (Monica, Sider) leave data in

~/Library/Safari/Extensions/

Copy

even after uninstalling. After removing any AI extension that had ‘access on all websites’ permission, clear the browser’s cookies to remove data it may have written across sites you visited.

Final cleanup checklist

Work through this in order. Each step builds on the previous one.

Audit current AI storage
```
du -sh ~/.ollama ~/.cache/lm-studio ~/jan ~/Library/Application\ Support/nomic.ai 2>/dev/null
```
Copy
Note the numbers. Run the same command again at the end to confirm what changed.

Stop and remove background services

Check for AI-related LaunchAgents:

ls ~/Library/LaunchAgents/ | grep -E ‘ollama|lmstudio|jan|gpt4all’

Copy

For each match, unload it before deleting:

launchctl unload ~/Library/LaunchAgents/com.ollama.ollama.plist

Copy

Then run the removal command:

rm ~/Library/LaunchAgents/com.ollama.ollama.plist

Copy

Remove app bundles
Move apps to Trash from
```
/Applications/
```
Copy
.

Delete model folders

Tool	Model folder to delete
Ollama	`~/.ollama/` Copy
LM Studio	`~/.lmstudio/models/` Copy
Jan	`~/jan/` Copy
GPT4All	`~/Library/Application Support/nomic.ai/` Copy

Clear app support and cache folders
Each tool leaves data in
```
~/Library/Application Support/
```
Copy
and
```
~/Library/Caches/
```
Copy
. App Cleaner & Uninstaller by Nektony finds all associated files automatically when you select an app, letting you remove any AI app completely.

Check MCP config files (Claude Desktop)
Open
```
~/.config/claude/claude_desktop_config.json
```
Copy
and check for API keys or tokens before deleting. Don’t skip this step when selling or transferring a Mac.

Clear browser data

Follow the browser-specific steps for any AI site you’ve used regularly.

Verify
Run the
```
du -sh
```
Copy
command from step 1 again. Then check System Settings → General → Storage, it won’t show model folders, but it gives a useful baseline for overall disk health.