ElfClean

ElfClean / Guides

Where Ollama and Hugging Face keep models on a Mac

A single local model is often 5 to 40 GB, and trying a few new ones a month can quietly take a hundred. Here's where each tool stores them on macOS, how to see what's using the space and how to remove the models you're done with.

Ollama

Ollama keeps every model you've pulled in ~/.ollama/models, unless you've set OLLAMA_MODELS to put them somewhere else. Inside are two folders: manifests, which names each model and tag, and blobs, which holds the actual weights.

Models built on the same base share blobs. That's why adding up the sizes ollama list shows can come to more than the folder really uses, and why deleting one model doesn't always free its whole size.

See what you have

ollama list                  # every model and tag, with its size
du -sh ~/.ollama/models      # what the folder really uses

Remove a model

ollama rm llama3.1:70b

Use the exact name and tag from ollama list. Ollama deletes the blobs no other model needs and keeps the shared ones. Don't delete files from blobs by hand: a blob that looks unused may belong to another model.

Hugging Face

Models, datasets and Spaces downloaded with transformers, diffusers, huggingface_hub or the Hugging Face CLI go into the hub cache at ~/.cache/huggingface/hub. HF_HOME or HF_HUB_CACHE moves it. Newer versions also keep a Xet cache beside it, in ~/.cache/huggingface/xet.

Each repository gets a folder named like models--meta-llama--Llama-3.1-8B, with blobs for the files, snapshots for each revision you've used and refs naming which revision is current. Old revisions keep their files until you remove them.

See what you have

huggingface-cli scan-cache

This lists every cached repository, its size and its revisions. Recent releases of huggingface_hub also install the shorter hf command; run hf cache --help to see its version of these commands.

Remove models

huggingface-cli delete-cache

It opens a checklist of cached repositories and revisions; tick the ones to delete and confirm. Deleting a whole models--… folder in Finder also works, as long as nothing is downloading into it at the time. A model you delete is downloaded again the next time code asks for it.

Other places models hide

ToolUsual location
LM Studio~/.lmstudio/models (older versions: ~/.cache/lm-studio/models)
OpenAI Whisper~/.cache/whisper
PyTorch Hub~/.cache/torch/hub
ComfyUIthe models folder of your ComfyUI install

Each app can be pointed somewhere else in its settings, so check there if the folder is empty.

How ElfClean does it

ElfClean lists every Ollama and Hugging Face model with the space it really uses. Files shared between models are counted once and kept until every model that uses them is selected, so removing one model never breaks another. It also finds Comfy Desktop downloads, and you can add model folders you keep elsewhere, even on another disk.

Models are never part of the Caches level, and nothing moves until you've reviewed it. What you remove goes to the Trash, where History can put it back.

More guides