Where AI models are stored on a Mac, and how to delete Ollama models and the rest

How we checked: every path and command here comes from the project's own documentation, release notes or source code, read on 6 and 7 October 2026, against the stable releases of 7 October (Ollama 0.40.0, huggingface_hub 2.1.1). Where a project does not document something, we say so. We make Strata, whose AI Data Sweeper reads most of these folders.

Where each AI tool keeps its models

Tool	Models are in	Delete with
Ollama	~/.ollama/models	ollama rm
LM Studio	~/.lmstudio/models, or ~/.cache/lm-studio/models on older installs	My Models, in the app
Hugging Face libraries	~/.cache/huggingface/hub	hf cache rm, hf cache prune
llama.cpp	The Hugging Face cache since March 2026; ~/Library/Caches/llama.cpp before	hf cache rm, or delete the file
Jan	~/Library/Application Support/Jan/data	Settings, then Delete All for an engine
GPT4All	~/Library/Application Support/nomic.ai/GPT4All	Not documented
ComfyUI Desktop	~/ComfyUI-Shared	Delete the file
Draw Things	~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models	Not documented
DiffusionBee	~/.diffusionbee (also holds your images)	Not documented
MacWhisper	~/Library/Application Support/MacWhisper/Models	Not documented
OpenAI Whisper	~/.cache/whisper	Delete the file
PyTorch, Keras	~/.cache/torch/hub, ~/.keras/models	Delete the file
Docker Model Runner	Docker's own store	docker model rm

Two warnings apply to the whole table. First, these are the defaults: if you set a custom folder in an app or with a variable such as OLLAMA_MODELS or HF_HOME, your models are there instead. Second, quit an app, or stop its server, before deleting its model files by hand, so nothing is using them while they go.

How to delete Ollama models

Ollama stores each model's weights as files named by their hash, such as sha256-3f1a…, in ~/.ollama/models/blobs, and a small manifest that lists them in ~/.ollama/models/manifests. Version 0.40.0, GitHub's current stable release when we checked on 7 October 2026, adds a manifests-v2 folder beside it. Its source shows that a model you pull gets an entry in both, while one you copy with ollama cp or build with an ordinary ollama create is listed only in the new folder. The commands below handle both. With Ollama running, list what you have, then remove what you no longer use:

ollama ls
ollama rm qwen3:8b

ollama ls shows each model's name, ID, size and when it was changed; ollama ps shows the ones loaded in memory, and ollama stop unloads one. ollama rm takes several names at once. It works through Ollama's server, so the app, or ollama serve, has to be running.

Why the space can be less than the size

Models share layers. Two tags of the same model, or a model and its variant, can use the same multi-gigabyte weights file. Ollama's source shows what ollama rm does: it deletes the manifest, then only the layers no other model references. Shared ones stay until the last model using them goes. Anything left over is usually cleared when the server next starts, unless OLLAMA_NOPRUNE is set; Ollama skips that clean-up while any manifest is unreadable.

Keeping models on another drive

Ollama reads OLLAMA_MODELS for a different folder, but the Mac app does not see a variable you export in your shell. Ollama's FAQ says to set it with launchctl, then quit and reopen the app:

launchctl setenv OLLAMA_MODELS "/Volumes/External/ollama-models"

Removing Ollama completely

Ollama's own macOS page lists what to remove: the app, its command at /usr/local/bin/ollama, and these folders. Two of the lines on that page put the ~ inside quotes, which zsh does not expand, so they quietly remove nothing. These are the working forms:

rm -rf ~/Library/Application\ Support/Ollama
rm -rf ~/Library/Saved\ Application\ State/com.electron.ollama.savedState
rm -rf ~/Library/Caches/com.electron.ollama ~/Library/Caches/ollama
rm -rf ~/Library/WebKit/com.electron.ollama
rm -rf ~/.ollama

The last line deletes every model. If you would rather be able to change your mind, move the same folders to the Trash instead: in Finder, or on macOS 15 or later with the trash command in place of rm -rf, as in trash ~/.ollama.

How to delete LM Studio models

LM Studio keeps models in ~/.lmstudio/models, one folder per publisher, then per model. That has been its home since version 0.3.6, released in January 2025, but its release notes say the change applied only to new installs and existing data was not moved. If you have used LM Studio since before then, look in ~/.cache/lm-studio/models too: you may have models in both.

LM Studio's documentation does not describe deleting a model. Guides do it from My Models in the app, and reports in LM Studio's bug tracker confirm that delete exists, some of them about it leaving empty folders behind, so check the models folder afterwards. Its command-line tool, lms, can list models with lms ls but has no delete command; a request for one has been open since June 2026.

How to clear the Hugging Face cache

Code that uses Hugging Face's libraries, from Transformers to diffusers, downloads models into ~/.cache/huggingface/hub, one folder per repository, such as models--openai--whisper-large-v3. Nothing is ever deleted automatically, and older revisions are kept. Hugging Face's hf command manages it:

hf cache ls
hf cache rm model/openai/whisper-large-v3 --dry-run
hf cache prune

hf cache ls lists what is cached and can filter by size or by when it was last used. hf cache rm removes a repository, a revision, or even one file. hf cache prune removes revisions nothing points to, unfinished downloads, and shared files no repository uses. Hugging Face's cache guide documents all three, and says a deleted file is downloaded again the next time it is needed.

Older guides give huggingface-cli delete-cache. That command was removed in huggingface_hub 1.0, in October 2025. Two neighbours of the hub folder are worth knowing: ~/.cache/huggingface/datasets holds prepared datasets, and ~/.cache/huggingface/xet holds upload data that Hugging Face's guide says can be deleted whole. And since March 2026, llama.cpp's -hf downloads also land in the hub folder, so they show up in hf cache ls.

Jan, ComfyUI, Whisper and the rest

• Jan keeps its data in ~/Library/Application Support/Jan/data, with models under llamacpp/models and mlx/models. Its settings have a Delete All for each engine, which leaves models you imported yourself alone.
• ComfyUI Desktop keeps its shared model library in ~/ComfyUI-Shared and its installs in ~/ComfyUI-Installs. ComfyUI's docs say to remove a model by deleting its file; the app offers no way to do it for you yet. Models you added by hand do not come back by themselves.
• OpenAI Whisper puts models in ~/.cache/whisper and downloads one again the next time it is asked for.
• MacWhisper keeps downloaded models in ~/Library/Application Support/MacWhisper/Models; its App Store version, Whisper Transcription, keeps them in its own sandbox container.
• Docker Model Runner manages its own: docker model list, docker model rm, and docker model purge to remove them all.
• GPT4All, Draw Things and DiffusionBee keep models in the folders in the table above. None of them documents how to delete a model, so quit the app first, and keep a note of what you removed in case you need it back.

Apple Intelligence's own models

Apple's page on getting Apple Intelligence says it needs up to 14 GB of storage on Macs with M3 or later and at least 12 GB of memory, and up to 8 GB on other Macs that support it. MacRumors measured more than 20 GB on one Mac running macOS 27.

What Apple does not offer is a way to remove them. Its macOS 27 guide has switches for individual features, such as Siri and message summaries, but no single Apple Intelligence switch, and no Apple page we found says that turning features off deletes the models. On our test Mac they sit in /System/Library/AssetsV2, which System Integrity Protection guards. Third-party tools that claim to remove them have appeared; Apple does not support them, and we have not tested them.

How Strata's AI Data Sweeper helps

Strata's AI Data Sweeper reads the usual folders of Ollama, LM Studio, Hugging Face, Jan, GPT4All, Whisper and other tools, plus any folder in ~/.cache that holds model files (not yet Comfy Desktop's ~/ComfyUI-Shared, Jan's MLX models or the App Store version of MacWhisper), and shows one card per tool with each model's size and a line on how it comes back, or for Ollama 0.40 and later, how it goes. For Ollama it reads the manifests, so models appear by name rather than by hash; before Ollama 0.40, a shared layer moves only when every model that uses it is chosen. For Hugging Face it counts the files a model keeps in the cache's shared blobs folder with that model, and moves such a file only with the last model that uses it. Whatever you choose goes to the Trash, where Put Back can return it. Apple's own models are shown so you can see the space, and never moved. Our step-by-step guide walks through the AI Data Sweeper, which is part of Strata Pro, included in every download's first 14 days. Strata's System Data room also gives AI models a part of their own, since Storage settings counts the models an app keeps in your Library folder as System Data, and those in a folder of their own, such as Ollama's, may count as Documents instead.

One exception to moving models to the Trash came with Strata 1.0.20. With Ollama 0.40 or later, Strata lists Ollama's models with their sizes and leaves removing them to Ollama: each row's Copy Command copies ollama rm and the model's name, to paste in Terminal while Ollama is running. Moving an Ollama 0.40 model's files would leave its manifests-v2 entry pointing at nothing, which stops Ollama clearing unused layers when it starts and stops ollama rm removing that model. If a model was removed that way with Strata 1.0.19, Put Back returns what Strata moved, or pulling the model again repairs the entry.

Frequently asked questions