Free space from Ollama and Hugging Face models

A single local model is often larger than every node_modules folder you own. Two commands show what you have:

ollama list
hf cache ls

Both print each model with its size. Removing one is as short: ollama rm and hf cache rm. What takes thought is which models you can get back, and how long that download will take.

Ollama

ollama list
du -sh ~/.ollama/models

ollama list shows each model with its tag, size and when it was last changed. Remove one by name:

ollama rm llama3.1:70b

Two things to know:

• Models share layers. Ollama stores models as blobs, and two tags of the same family can use the same blob. Removing one model frees only the blobs no other model needs, so the space you get back can be less than the size in the list.
• Do not delete blobs by hand. The files in ~/.ollama/models/blobs are named by checksum and referenced from manifests. Remove a blob directly and the models that use it break.

Hugging Face

Everything that transformers, diffusers or huggingface_hub downloads goes to ~/.cache/huggingface/hub, one folder per repository:

hf cache ls

The list has one line per repository with its size and when it was last accessed. It takes filters, so this shows the large ones you have not touched for a month:

hf cache ls --filter "size>1GB" --filter "accessed>30d"

Remove a repository, with a preview first:

hf cache rm model/bert-base-cased --dry-run
hf cache rm model/bert-base-cased

And remove leftovers that nothing points to: old revisions of models that were updated, and files from downloads that were interrupted:

hf cache prune --dry-run
hf cache prune

hf cache prune is the safe first step. It removes nothing that a current model reference needs.

If hf is not found, your huggingface_hub is older. The same job there is huggingface-cli scan-cache and huggingface-cli delete-cache.

Datasets are separate. The datasets library keeps its processed copies in ~/.cache/huggingface/datasets, and they can be larger than the models:

du -sh ~/.cache/huggingface/*

LM Studio and stray model files

LM Studio keeps models in ~/.lmstudio/models, in older versions in ~/.cache/lm-studio/models. Delete them in the app, under My Models, so that its list stays right.

Other tools put model files wherever you told them to. Find every large model file in your home folder, wherever it is:

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

It is common to find the same model three times: once in Ollama, once in the Hugging Face cache, once in a downloads folder.

What to keep

• Models you run every week. The download is not worth the space.
• Anything you made. Fine-tunes, merges, quantized conversions. Move them to an external disk before a cleanup.
• Models that are no longer published. If the repository was removed, your copy is the last one.

The Python side of an AI setup has its own weight. Every virtualenv with PyTorch in it holds gigabytes of libraries. See pip, conda and virtualenvs.

Move them instead of deleting them

If the models are worth keeping and the internal disk is the problem, both tools can keep their store on an external disk.

For Ollama, the folder is set by the OLLAMA_MODELS variable. The Mac app reads its environment from launchctl, so set it there, then quit and reopen Ollama:

launchctl setenv OLLAMA_MODELS /Volumes/External/ollama-models

For Hugging Face, HF_HOME moves the whole cache. Put it in your shell profile:

export HF_HOME=/Volumes/External/huggingface

In both cases, move the existing folder to the new place first, so that nothing has to be downloaded again. LM Studio has the same option in its settings, as the models folder.

The trade is plain: models load more slowly from an external disk, and they are not there when the disk is not plugged in.