A local model is one of the largest files most people will ever keep on a Mac: a few gigabytes for a small one, tens of gigabytes for a large one. Try a handful and they quickly outgrow everything else. They also stay until you remove them, even the ones you tried once.
Each tool keeps its models in its own place and its own way, which decides how to remove them safely.
Ollama
Ollama keeps every model you pull or create in ~/, but not as one file per model. Each model is a small manifest plus the layers it's made of, stored in a blobs folder under names that are long hashes. Models built on the same base share those layers.
That's why you shouldn't remove files there in Finder. You can't tell which blob belongs to which model, and a layer you remove may be one another model still needs.
Ollama itself removes a model along with only the layers no other model uses. Ollama's app has no button for this; its command-line tool does it, one model at a time. A removed model downloads again the next time you pull or run it. Models you created yourself from a Modelfile need to be created again, so keep the Modelfile.
If models are filling your startup disk, Ollama's app can keep them on another disk instead: open its Settings and change Model location.
LM Studio
LM Studio keeps models in ~/, in a folder for each publisher and model. Earlier versions used ~/, so check both.
Remove them in LM Studio: open My Models, click the menu next to a model you no longer need, and choose Delete. A deleted model can be downloaded again from the Discover tab. Models you imported from elsewhere need to be imported again, so keep the original file if you want it back.
Hugging Face
Python libraries such as Transformers and Diffusers, Hugging Face's own command-line tool and apps such as llama.cpp download models and datasets into ~/, in a folder for each repository, named like models--publisher--name. New revisions of a model are stored next to the old ones, and nothing is removed automatically.
Moving a repository's folder to the Trash removes that model and nothing else. Programs download it again when they need it, which can take a long time, and gated models need you to be signed in to Hugging Face first. Some newer versions of Hugging Face's library keep the files themselves in a shared blobs folder next to the repositories; then a model's folder frees little on its own, and Hugging Face's own tool clears what no model uses.
Chrome's on-device model
On eligible Macs, Chrome downloads Gemini Nano, a model of several gigabytes, so that some of its features can run on your Mac. It lives in Chrome's own folder, updates in the background, and Chrome removes it by itself when free space runs low.
To remove it and stop it downloading again, use Chrome's setting: click the More menu, choose Settings, then AI Innovations, and turn off On-device AI.
Apple Intelligence
Apple Intelligence's models belong to macOS, in folders it protects, and Storage settings counts them under macOS. There's no supported way to delete them. If you don't use Apple Intelligence, turning it off in System Settings › Apple Intelligence & Siri lets macOS manage that space, though it may not come back right away.
How the app does it
Select a model folder in the Overview's map or list, and the inspector says what it is, who made it and how it's removed. The Cleanup page lists the model folders the catalog knows, by level. Models are your own choice to keep, so nothing there starts checked:
- LM Studio's models folder and Hugging Face's cache are Review first. Checking one moves the whole folder, every model in it, to the Trash, where you can put it back. To remove a single model, use the steps above.
- Ollama's models are explained, but the app never removes them: only Ollama knows which layers each model still needs, so remove models in Ollama, as above.
- Chrome's on-device model is explained with Chrome's own setting, which removes it and stops it downloading again.
Seeing what your models take is free. Moving them to the Trash is PRO.
Safe to delete?
The entries this guide is about, in the words the app shows. Each has its own page with its sources.
~/
.ollama/ models Language models you've pulled or created with Ollama, stored as shared layers.
Review first · Explained only: never removed
~/
.lmstudio/ models Language models downloaded or imported in LM Studio, in a folder for each publisher and model.
Review first · Moves it to the Trash
~/
.cache/ huggingface/ hub Models and datasets downloaded from Hugging Face, shared by every library and tool that uses them.
Review first · Moves it to the Trash
~/
Library/ Application Support/ Google/ {Chrome,Chrome Beta,Chrome Dev,Chrome Canary}/ OptGuideOnDeviceModel The Gemini Nano language model Chrome downloads in the background to run AI features on your Mac.
Usually safe · Explained only: the app shows you how