# Where Ollama, LM Studio and Hugging Face store models on a Mac, and how to remove them

> Find every local AI model on your Mac: Ollama, LM Studio, the Hugging Face cache, ComfyUI and more. Remove them with each tool’s own command, or move them.

- URL: https://moreroom.app/learn/local-ai-models-storage
- Published: 2026-10-17
- Updated: 2026-10-17
- Publisher: MoreRoom (Yando Pte. Ltd.) — https://moreroom.app

## TL;DR

- Ollama stores models in ~/.ollama/models, LM Studio in ~/.lmstudio/models and Hugging Face tools in ~/.cache/huggingface/hub.
- Each tool keeps its own copy, so the same model downloaded in three apps takes three times the space.
- Remove models with each tool’s own command, such as ollama rm or hf cache rm, because files inside the caches are shared between models.
- Models are typically 2–70 GB each and can be downloaded again, but your own fine-tunes and generated images cannot.
- To keep models on an external SSD, use the official settings: Ollama’s model location, LM Studio’s models directory or HF_HOME.

Local AI tools keep their [model weights](https://moreroom.app/glossary#model-weights) in hidden folders in your home folder: **Ollama in `~/.ollama/models`**, **LM Studio in `~/.lmstudio/models`** and **Hugging Face tools in `~/.cache/huggingface/hub`**. Each tool keeps its own copy, and a single model is typically anywhere from 2 GB to 70 GB. Remove models through each tool’s own command or app, not by deleting files inside the caches, because those files are shared between models.

## Where each tool keeps its models

| Tool | Default location | How to remove |
|---|---|---|
| Ollama | `~/.ollama/models` | `ollama rm <model>` |
| LM Studio | `~/.lmstudio/models` | My Models in the app |
| Hugging Face (transformers, diffusers, MLX) | `~/.cache/huggingface/hub` | `hf cache rm` |
| ComfyUI Desktop | `~/Documents/ComfyUI/models` | Finder, by model type |
| Draw Things | `~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models` | In the app |
| PyTorch Hub | `~/.cache/torch` | Finder |
| OpenAI Whisper | `~/.cache/whisper` | Finder |

Older LM Studio versions used `~/.cache/lm-studio/models`, and a manual ComfyUI install keeps its models in the `models` folder of wherever you installed it, often `~/ComfyUI/models`.

![Map of where Ollama, LM Studio, Hugging Face and other AI tools store models on a Mac](https://moreroom.app/learn/img/local-ai-models-storage-map.svg)

## Ollama

Ollama’s [FAQ](https://docs.ollama.com/faq) gives `~/.ollama/models` as the macOS default. Inside, `blobs` holds the actual data and `manifests` records which blobs make up each model. Several models built on the same base can share blobs, so deleting files by hand can break other models. Use the CLI instead:

```sh
ollama list
ollama rm llama3.2
```

`ollama list` (also `ollama ls`) shows every model with its size. `ollama rm` removes the model and any blobs no other model needs. It is permanent; `ollama pull` downloads the model again. If a model is loaded in memory, `ollama ps` shows it and `ollama stop <model>` unloads it, but neither frees disk space.

## LM Studio

Current LM Studio versions keep models in `~/.lmstudio/models`, organised as publisher, then model, then the `.gguf` or MLX files. Open the **My Models** tab to see each model with its size and delete the ones you no longer use. The same tab shows, and lets you change, the models directory.

In Terminal, `lms ls` lists downloaded models. If you prefer Finder, quit LM Studio first, then move a model’s folder to the Trash; LM Studio no longer lists it.

## The Hugging Face cache

Python libraries such as transformers and diffusers, MLX tools that download from the mlx-community organisation, and many Mac apps built on them all download through `huggingface_hub` into one shared cache. Its location is `~/.cache/huggingface/hub` unless you set `HF_HOME` or `HF_HUB_CACHE`.

As of October 2026 the official CLI is `hf`, which replaced the older `huggingface-cli`. The [cache management guide](https://huggingface.co/docs/huggingface_hub/guides/manage-cache) documents three commands:

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

- `hf cache ls` lists each cached repository with its size, last access and references. Filters such as `--filter "accessed>30d"` or `--filter "size>1GB"` narrow the list.
- `hf cache rm` deletes whole repositories, specific revisions or, with an `hf://` path, a single file such as one GGUF quantisation. `--dry-run` shows what would go first. Deletion is permanent.
- `hf cache prune` removes revisions no branch or tag points to, leftover `.incomplete` partial downloads and shared files nothing uses any more.

If your installation only has `huggingface-cli`, its older `scan-cache` and `delete-cache` commands do the same job; upgrading `huggingface_hub` gives you `hf`.

![How the Hugging Face cache links refs, snapshots and blobs](https://moreroom.app/learn/img/local-ai-models-storage-hf-cache.svg)

Why not just delete a `models--…` folder? Inside each one, `snapshots` contains symbolic links that point to the real files in `blobs`. Recent `huggingface_hub` versions can also keep files downloaded through Xet in a shared `blobs` folder at the top of the cache, used by several repositories at once. Deleting a model folder by hand can then free less than expected; `hf cache prune` afterwards removes shared files that no model uses.

Two neighbours of the hub cache also grow: `~/.cache/huggingface/datasets` holds datasets processed by the datasets library, and `~/.cache/huggingface/xet` holds data for Hugging Face’s Xet storage system. Current versions use it mainly for uploads; older versions also kept a download chunk cache there.

## ComfyUI, Draw Things and other apps

- **ComfyUI** sorts models by type in folders such as `checkpoints`, `loras` and `vae`. Most can be downloaded again, but some community models are hard to find later, so archive rather than delete what you are unsure about. Its `output` folder holds the images you generated: that is your work, not cache.
- **Draw Things** keeps models inside its app container. Delete them from the app’s model manager rather than in Finder.
- **PyTorch Hub** (`~/.cache/torch`) and **openai-whisper** (`~/.cache/whisper`) download weights on first use and fetch them again when a script needs them.

## Find stray model files

Model files also end up in Downloads, project folders and old experiments. This command lists `.gguf`, `.safetensors` and `.ckpt` files larger than 1 GB anywhere in your home folder, largest last:

```sh
find ~ -type f \( -name "*.gguf" -o -name "*.safetensors" -o -name "*.ckpt" \) -size +1G -exec du -h {} + 2>/dev/null | sort -h
```

It only reads; nothing is deleted. Ollama and Hugging Face store their blobs under hash names without extensions, so the results show files outside those caches. Terminal may ask for access to folders such as Documents or Desktop. Before deleting a result, check whether it is your own fine-tune or a model you cannot download again.

## Move models to an external SSD

Use each tool’s own setting rather than moving folders and hoping it follows:

- **Ollama**: recent versions of the Ollama app have a **Model location** setting. The documented alternative is the `OLLAMA_MODELS` environment variable; for the Mac app, set it with `launchctl setenv OLLAMA_MODELS "/Volumes/AI/ollama"` and restart Ollama. That setting does not survive a restart of the Mac. Existing models are not moved for you: quit Ollama and move the contents of `~/.ollama/models` yourself, or download them again.
- **LM Studio**: change the models directory in **My Models**.
- **Hugging Face**: add `export HF_HOME="/Volumes/AI/huggingface"` to `~/.zshrc`, or set `HF_HUB_CACHE` to move only the hub cache. See the [environment variables reference](https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables). Apps opened from the Finder do not read `~/.zshrc`.
- **ComfyUI**: add folders in the Desktop app’s **Storage** settings or in `extra_model_paths.yaml`, as described in [ComfyUI’s model docs](https://docs.comfy.org/development/core-concepts/models).

Format the drive as APFS: the Hugging Face cache relies on symbolic links, which exFAT does not support. When the drive is not connected, tools fail to load models or start downloading them again to the default location.

## How MoreRoom helps

MoreRoom’s free version recognises the folders of Ollama, LM Studio, Hugging Face, PyTorch, Whisper, ComfyUI, Draw Things and other AI tools, shows their real size, and marks them “Review first”: they can be downloaded again, but only you know which ones you still use. Ollama models are removed with `ollama rm`, a permanent command labelled as such. Hugging Face models go to the Trash one at a time, with Undo; with a recent `huggingface_hub`, run `hf cache prune` afterwards to release shared files. For LM Studio and Draw Things, MoreRoom tells you where to delete them in the app. Large model files found elsewhere and ComfyUI outputs are marked to keep; offloading them to an external drive from MoreRoom is coming soon.

Pro adds the Collector, to clean several items in one go, and Goal mode.

## FAQ

### Where does Ollama store models on a Mac?

In ~/.ollama/models, split into a blobs folder with the model data and a manifests folder that lists which blobs make up each model. You can change the location in the Ollama app’s settings or with the OLLAMA_MODELS environment variable.

### Is it safe to delete ~/.cache/huggingface?

Usually, yes: it is a download cache, and models download again the next time a script or app needs them. Check first that no project stores irreplaceable files there, and prefer hf cache rm for individual models.

### Why did deleting a model folder free less space than expected?

The tools share files between models. Ollama shares layers between models, and recent Hugging Face versions can store downloaded files once in a shared blobs folder. Removing through ollama rm or hf cache rm followed by hf cache prune frees what is truly unused.

### Do Ollama and LM Studio share downloaded models?

No. Each app downloads and stores its own copy in its own folder, so a model you use in both takes space twice.

### Can I keep AI models on an external drive?

Yes. Set the model location in Ollama, change the models directory in LM Studio’s My Models tab, or set HF_HOME for Hugging Face tools. Format the drive as APFS, and expect the tools to fail or download again when it is not connected.