Delete Ollama, Hugging Face and LM Studio models on a Mac

How much space are your models using?

One line per tool. Each only reads:

# Ollama: every model with its size
$ ollama list

# Hugging Face: every cached model and dataset with its size
$ uvx --from huggingface_hub hf cache ls

# LM Studio: the models folder (older versions used ~/.cache/lm-studio/models)
$ du -sh ~/.lmstudio/models/*/* 2>/dev/null

uvx runs the Hugging Face CLI without installing it. If you already have it (pip install -U huggingface_hub), drop the uvx --from huggingface_hub part and run hf cache ls.

Ollama

Ollama stores everything under ~/.ollama/models:

• manifests/ holds one small file per model and tag, such as registry.ollama.ai/library/llama3.1/8b.
• blobs/ holds the actual weights, named by their SHA-256 hash.

Models built on the same base share blobs. That's why adding up the folder by hand and comparing it with ollama list can give different numbers, and why you shouldn't delete files in blobs/ yourself.

$ ollama list                # name, ID, size, modified
$ ollama rm llama3.1:8b      # remove one model
$ ollama rm qwen3:4b gemma3  # or several at once

ollama rm deletes the model's manifest and every blob no other model needs. It doesn't use the Trash, so the only way back is ollama pull.

If you moved your models with the OLLAMA_MODELS environment variable, they're in that folder instead. ollama rm still finds them, as long as the Ollama server was started with the same setting.

Hugging Face

transformers, diffusers, sentence-transformers, mlx-lm and most other Python AI libraries download models into the Hugging Face cache:

$ ~/.cache/huggingface/hub

HF_HOME or HF_HUB_CACHE moves it. Each model has a folder named models--org--name, and datasets use datasets--org--name.

Don't judge those folders by their size in Finder. Recent versions of huggingface_hub keep the large weight files in a shared blobs folder at the top of the cache and link to them from each model's folder. On my Mac the folder for mlx-community/whisper-large-v3-turbo is 20 KB, while the model really takes 1.6 GB. Delete that folder in Finder and the 1.6 GB stays.

The hf CLI understands the layout:

# list with sizes, last used and last modified
$ hf cache ls

# see what a delete would remove, then do it
$ hf cache rm model/mlx-community/whisper-large-v3-turbo --dry-run
$ hf cache rm model/mlx-community/whisper-large-v3-turbo

# old revisions and half-finished downloads only
$ hf cache prune

hf cache rm asks before it deletes (-y skips the question) and accepts several repos at once. hf cache prune keeps the models you use and removes only revisions no branch points to and downloads that never finished, so it's the safe one to run first.

LM Studio

LM Studio keeps each model in its own folder, grouped by publisher:

$ ~/.lmstudio/models/<publisher>/<model>

Older versions used ~/.cache/lm-studio/models, and you can change the folder in LM Studio's settings. Delete models from LM Studio's own model list, so the app knows they're gone. If you remove a folder by hand, quit LM Studio first.

Models you downloaded through LM Studio come back from its Discover tab. A GGUF file you copied in yourself only comes back if you still have it somewhere else.

Shedi lists every Ollama and LM Studio model with its sizeOne row per model, with what it is and how to get it back. Ollama models are counted the way Ollama counts them, so shared files stay. Selected models go to the Trash first.

Get Shedi

Other AI caches worth a look

These are the default folders. All of them download again when needed:

Each Python virtual environment also has its own copy of PyTorch and friends. The developer cache reference covers project .venv folders, along with the pip and uv caches.

A sensible routine

$ ollama list
$ hf cache ls
# remove what you haven't used this month
$ ollama rm <model>
$ hf cache rm <repo>
$ hf cache prune

Models are usually the biggest item, but rarely the only one. To see them next to Docker, Xcode and package caches, run diskreceipt, a free Terminal tool that prints every developer cache with its size, or open Shedi.

Sources

• Ollama docs: FAQ, "Where are models stored?" and OLLAMA_MODELS
• Ollama CLI reference
• Hugging Face Hub docs: Manage the cache
• Hugging Face Hub docs: Command line interface
• Hugging Face Hub docs: Environment variables (HF_HOME, HF_HUB_CACHE)
• LM Studio docs
• PyTorch docs: torch.hub
• Checked on this Mac on 28 October 2026: ollama rm --help (Ollama 0.32.1), and hf cache ls, hf cache rm --dry-run and hf cache --help (huggingface_hub 2.3.0).