How to Remove Unused Ollama Models on Mac

Where Ollama keeps its models

Everything lives under ~/.ollama. The weights themselves are in ~/.ollama/models, and that single folder is almost always the reason Ollama shows up on a disk-usage scan. Check it:

du -sh ~/.ollama/models

Ollama doesn't store one file per model. It stores content-addressed blobs — each layer of a model named after its SHA-256 hash — plus small JSON manifests that say which blobs make up each model:

The blobs are labelled Warning, not Safe — not because deleting them breaks anything, but because each one is a multi-gigabyte download to get back. Treat them like luggage you might re-pack, not trash.

See what you actually have

The right way to inspect models is Ollama's own command — it reads the manifests and shows real names and sizes:

ollama list

You'll get something like:

NAME                ID              SIZE      MODIFIED
llama3.1:8b         42182419e950    4.9 GB    3 weeks ago
qwen2.5-coder:32b   4bd6cbf2d094    19 GB     2 months ago
deepseek-r1:7b      0a8c26691023    4.7 GB    5 days ago

The MODIFIED column is your cleanup signal: anything you last touched months ago is a strong candidate. To confirm what the blobs weigh on disk:

du -sh ~/.ollama/models/blobs

Remove a model you don't use

Always remove through Ollama so the manifest and its blobs are cleaned together:

ollama rm qwen2.5-coder:32b

Remove several at once:

ollama rm llama2:13b vicuna:7b orca-mini:3b

When you want one of them back later, it's a single command — Ollama re-pulls it from the registry:

ollama pull llama3.1:8b

The nuance most guides miss: shared layers

Because blobs are content-addressed, two models built on the same base share the identical layer on disk — it's stored once. That has a consequence people find surprising: ollama rm only deletes layers no other model still references.

So if ollama list says a model is 4.9 GB but removing it frees only 1.2 GB, the rest was shared with a model you kept. Nothing went wrong — you simply weren't carrying 4.9 GB of unique data. The way to actually reclaim a lot is to remove a family of related models, or the one large standalone model that shares nothing.

Optional: move the model store to another disk

If you collect large models, point Ollama at an external SSD instead of cleaning constantly. Set OLLAMA_MODELS and restart Ollama:

launchctl setenv OLLAMA_MODELS /Volumes/AI/ollama-models

New pulls land there. Inference is fastest from internal NVMe, so this is a capacity trade, not a speed win — but it keeps tens of gigabytes off your boot drive.

FAQ

Is it safe to delete Ollama models?

Yes — nothing breaks. They re-download with ollama pull; the only cost is bandwidth. Just remove them with ollama rm, never by deleting blobs by hand.

How big do Ollama models get?

A quantized 7–8B model is ~4–5 GB; a 32B is ~19 GB; 70B models run 40 GB and up. A few experiments and the ~/.ollama/models folder is easily 50–100 GB.

Does ollama rm free the whole listed size?

Only the layers nothing else uses. Models that share a base layer keep that shared blob until the last model using it is removed.

Can I just delete the whole ~/.ollama folder?

You can, and Ollama rebuilds it on next launch — but you also lose your keypair and every downloaded model at once. Prefer ollama rm for the models you actually want gone.