Every tool keeps its weights somewhere different, none of them tell you where, and together they are the fastest-growing thing on a developer's disk. Here is the full list, per platform.
Free to download and scan. Cleaning needs a licence: one payment per computer, Mac or Windows.
Mainstream Mac cleaners were written before local models existed, and they work from a fixed list of known cache folders. ~/.ollama is not on it. So a tool will confidently report your Mac is clean while 120 GB of weights sit in a hidden folder in your home directory — and because they are dot-folders, the Finder does not show them either.
Bytesweep macOS ~/.ollama/models
Linux /usr/share/ollama/.ollama/models
Windows %USERPROFILE%\.ollama\modelsList and remove with Ollama itself, so its manifest stays consistent:
ollama list
ollama rm llama3:70bMove the store to an external drive by setting OLLAMA_MODELS before starting the server.
macOS ~/.lmstudio/models
Linux ~/.lmstudio/models
Windows %USERPROFILE%\.lmstudio\modelsOlder versions used ~/.cache/lm-studio/models; if you have upgraded through several releases, check both. The directory is configurable in the app's My Models tab, which is the safe way to relocate it.
macOS/Linux ~/.cache/huggingface/hub
Windows %USERPROFILE%\.cache\huggingface\hubUsually the largest and the messiest, because every library that touches the Hub writes here. Inspect and prune interactively with the official tool:
pip install "huggingface_hub[cli]"
huggingface-cli scan-cache
huggingface-cli delete-cacheRelocate it with HF_HOME. Note that the hub cache uses symlinks into a blobs folder — copying it with a tool that follows symlinks will double its size.
~/.cache/torch/hub
~/.cache/torch/kernelsPretrained checkpoints downloaded by torch.hub. Relocate with TORCH_HOME.
macOS ~/Library/Application Support/nomic.ai/GPT4All
Linux ~/.local/share/nomic.ai/GPT4All
Windows %LOCALAPPDATA%\nomic.ai\GPT4All~/.cache/whisper
./models (inside a whisper.cpp checkout)~/.cache/torch/sentence_transformers
~/.keras/models
~/.cache/gguf
~/.cache/vllmEverything else on a developer's disk regenerates in seconds. A 40 GB model is a re-download over your connection, and a quantisation you produced locally or a model you fine-tuned yourself does not come back at all. These folders deserve review, not a sweep.
Ollama and Hugging Face both keep a manifest alongside the blobs. Delete the files yourself and the manifest still lists them, which shows up later as confusing errors. ollama rm and huggingface-cli delete-cache exist for this.
Every tool above supports relocating its store through an environment variable or a setting. If you have an external drive, moving the store beats repeatedly deleting and re-pulling the same weights.
Set the OLLAMA_MODELS environment variable to the directory you want, then restart the Ollama server. On macOS, set it for the app's launch environment rather than only in your shell, or the desktop app will not see it.
Safe in the sense that nothing breaks — anything missing is re-downloaded on next use. Expensive in the sense that "next use" may mean pulling 30 GB again. Use huggingface-cli scan-cache to see what is there and delete selectively.
Most of them start with a dot, which macOS hides by default. Press Cmd-Shift-. in a Finder window to show them, or open the path directly with Go → Go to Folder.
Yes, and it is the better answer once you pass a hundred gigabytes. Ollama uses OLLAMA_MODELS, Hugging Face uses HF_HOME, PyTorch uses TORCH_HOME, and LM Studio has a setting in the app. Expect slower first loads over USB.
Usually partial downloads and superseded revisions. The Hugging Face hub cache keeps every revision you have ever pulled of a repository, not just the current one, which is exactly what scan-cache is designed to show you.
Shorter pieces that go deeper on one part of this guide.