How to delete Ollama models and free disk space

Last updated 17 October 2026

Local models are big: from a few gigabytes for a small model to 40 GB or more for a 70B one. After a few weeks of trying models, it's easy to have 100 GB sitting in a folder you never look at. Here's how to see what you have and remove what you don't use.

See what's installed

ollama list

This shows each model with its size and when it was last modified. ollama ps shows which models are loaded in memory right now.

Remove a model

ollama rm llama3:70b

Use the exact name and tag from ollama list. If you want it back later, ollama pull llama3:70b downloads it again.

Why not just delete the files?

On a Mac, Ollama keeps models in ~/.ollama/models, split into manifests and blobs. Blobs are named by their SHA-256 hash and can be shared between models, so deleting one by hand can break another model that uses the same layer. ollama rm knows which blobs are still needed and removes only the rest.

If you've set the OLLAMA_MODELS environment variable, models live in that folder instead.

Other places models hide

LM Studio keeps downloaded models in ~/.lmstudio/models (older versions used ~/.cache/lm-studio). You can delete models from LM Studio's My Models screen.
Hugging Face downloads go to ~/.cache/huggingface/hub. If you have the huggingface_hub command-line tool, huggingface-cli scan-cache shows what's there and huggingface-cli delete-cache lets you pick what to remove.
MLX models downloaded through Hugging Face use the same cache.

Keep the ones you use, move the ones you might

If you have a NAS or an external SSD, a model you use once a month doesn't need to live on your internal disk. Moving the whole ~/.ollama/models folder to an external drive and pointing OLLAMA_MODELS at it works, as long as the drive is connected when you run Ollama.

How DiskCity handles models

DiskCity is in development. Its AI models rule reads Ollama's manifests (metadata only) to name each model's blobs, marks models you haven't used in 60 days, and removes them with ollama rm rather than deleting files, logging the ollama pull command that brings each one back. LM Studio and Hugging Face models are recognised too. In Pro, the Harbour can move models to your own NAS or SSD instead. Join the waitlist to hear when it launches.