WTMemory guide

Where Ollama stores models on Mac, and how to delete them

Updated 13 October 2026

On a Mac, Ollama stores downloaded models in ~/.ollama/models, and you delete one with ollama rm <model>, which frees its space immediately (it does not go to the Trash). Other local AI tools keep their own copies: Hugging Face in ~/.cache/huggingface/hub, LM Studio in ~/.lmstudio/models, and ComfyUI in the models folder of its install.

A handful of models can easily use 50 to 200 GB. This guide shows how to measure each folder, remove models safely, and move them to an external drive. If you would rather see all of these locations in one window, WTMemory inspects them for free from the menu bar.

Where does Ollama store models on Mac?

By default, everything lives under ~/.ollama/models:

If you set the OLLAMA_MODELS environment variable, Ollama uses that folder instead. Check both:

echo $OLLAMA_MODELS
launchctl getenv OLLAMA_MODELS
du -sh ~/.ollama/models

To see what you have installed and how big each model is:

ollama list      # name, ID, size, last modified
ollama ps        # models currently loaded in memory

Models built on the same base can share blobs, so the sizes in ollama list may add up to more than du reports. du shows the real disk usage.

How do I delete Ollama models?

Use Ollama's own command so the manifests and blobs stay consistent:

ollama list
ollama stop llama3.1:8b   # unload it first if ollama ps shows it running
ollama rm llama3.1:8b

ollama rm removes the manifest and deletes any blobs no other model still uses. The files are deleted directly, not moved to the Trash, so the space comes back at once. You can download the model again later with ollama pull.

Avoid deleting files inside blobs by hand. A missing blob breaks every model that references it, and you will get errors until you pull those models again. Custom models you created with ollama create from your own Modelfile and weights cannot be pulled again, so keep a copy of the source files before removing them.

To remove Ollama completely, quit the app, delete /Applications/Ollama.app, and then remove ~/.ollama. That folder also contains your Ollama key pair, so only delete it when you are done with Ollama.

How big are local LLM models?

Size depends mostly on parameter count and quantization. These are rough figures for common 4-bit (Q4) GGUF builds; exact sizes vary by model family and quantization type:

Model sizeApproximate disk size at Q4
1B to 3Babout 1 to 2 GB
7B to 8Babout 4 to 5 GB
13B to 14Babout 8 to 9 GB
30B to 34Babout 18 to 20 GB
70Babout 40 GB

Higher-precision builds (Q8, FP16) are roughly two to four times larger. Image models are big too: a Stable Diffusion XL checkpoint is around 6 to 7 GB, and full-precision Flux checkpoints can exceed 20 GB.

How do I move Ollama models to an external drive?

  1. Quit Ollama from its menu bar icon so nothing is writing to the folder.
  2. Copy the models to the drive, for example: rsync -a ~/.ollama/models/ /Volumes/External/ollama-models/
  3. Point Ollama at the new location. For the Mac app, run launchctl setenv OLLAMA_MODELS /Volumes/External/ollama-models. If you start the server yourself with ollama serve, add export OLLAMA_MODELS=/Volumes/External/ollama-models to your ~/.zshrc instead.
  4. Start Ollama again and run ollama list. Once your models appear, remove the old copy in ~/.ollama/models.

Two caveats. A launchctl setenv value does not survive a restart, so you need to set it again after rebooting (or use a login item that runs the command). And the drive must be mounted before Ollama starts, or Ollama will fail to find your models. Recent versions of the Ollama app may also offer a model location option in their settings; check yours before using the environment variable.

Where is the Hugging Face cache on Mac?

Models downloaded through the huggingface_hub library, transformers, diffusers, mlx-lm and many other Python tools land in ~/.cache/huggingface/hub. Each repository gets a folder such as models--meta-llama--Llama-3.1-8B, with blobs, snapshots and refs inside. The HF_HOME or HF_HUB_CACHE variables move it elsewhere.

du -sh ~/.cache/huggingface/hub
du -sh ~/.cache/huggingface/hub/* | sort -h

The Hugging Face CLI can list and remove cached repos for you:

huggingface-cli scan-cache     # list repos, sizes, last used
huggingface-cli delete-cache   # pick revisions to delete interactively

Newer releases of huggingface_hub ship an hf command that replaces huggingface-cli. Run hf cache --help to see the cache subcommands your version provides.

Do not delete ~/.cache/huggingface as a whole. The parent folder can contain your access token (token and stored_tokens). Deleting a single models--... folder inside hub is fine; the files download again the next time a script needs them.

Where does LM Studio store models?

Current versions of LM Studio keep models in ~/.lmstudio/models, organized by publisher and model name. Older versions used ~/.cache/lm-studio/models, which can linger after an upgrade and quietly hold a duplicate set of files.

du -sh ~/.lmstudio/models ~/.cache/lm-studio/models 2>/dev/null

The safest way to delete a model is from the My Models section inside LM Studio, which is also where you can change the models directory. If you have the lms command line tool, lms ls lists what is downloaded. Before removing the legacy folder, open LM Studio and confirm the models you still want appear from the new location.

How much space do ComfyUI models use?

ComfyUI stores models inside its own install, in ComfyUI/models, split into folders like checkpoints, loras, vae, clip, unet or diffusion_models, controlnet and upscale_models. Common install locations are ~/ComfyUI and ~/Documents/ComfyUI.

du -sh ~/ComfyUI/models/* ~/Documents/ComfyUI/models/* 2>/dev/null | sort -h

Checkpoints are usually the largest items. Be careful with LoRAs and merged models you trained or downloaded from sites that may remove them later: they may not be available again. If you use the same models in several tools, an extra_model_paths.yaml file lets ComfyUI read them from a shared folder instead of keeping duplicates.

Which AI model files are safe to delete?

Is there a faster way to see all of this at once?

WTMemory is a native menu-bar app for macOS 12 and later. Its free AI Tools view shows supported Ollama, Hugging Face Hub cache, LM Studio (current and legacy) and ComfyUI model locations with their estimated sizes, next to recognized AI processes and the memory and ports they use. The free Disk Explorer covers custom model folders. With Pro ($9 once, up to three Macs, 7-day trial) you can select individual locations, read the consequences, and move them to the Trash after review. Nothing is deleted automatically, and the Hugging Face parent folder is never touched. It does not decide which models you no longer need; that stays your call.

Download WTMemory

Frequently asked questions

Where are Ollama models stored on a Mac?

In ~/.ollama/models by default. The weights are in the blobs subfolder and the name mappings are in manifests. If OLLAMA_MODELS is set, Ollama uses that folder instead.

Does ollama rm free up disk space?

Yes. ollama rm deletes the model's manifest and any blobs no other model uses, and the space is freed immediately because the files are not moved to the Trash.

Can I delete the Hugging Face cache folder?

Delete individual repos inside ~/.cache/huggingface/hub or use the Hugging Face CLI cache commands. Do not delete ~/.cache/huggingface wholesale, because it can contain your access token.

How do I change where Ollama saves models on Mac?

Quit Ollama, copy the models folder to the new location, run launchctl setenv OLLAMA_MODELS with the new path, then restart Ollama. The launchctl setting has to be applied again after a reboot.

How big is a 7B or 8B model?

A 4-bit quantized 7B or 8B model is roughly 4 to 5 GB. A 4-bit 70B model is roughly 40 GB. Higher-precision versions are several times larger.

Related guides