Guides · Updated

Keeping local models up to date and tidy (disk space, updates, removing old models)

See what's using your disk, remove models you don't need, update Ollama and your models without surprises, and clear out the hidden caches that other AI tools leave behind.

Not sure which model to pull? Use the GPU model picker: enter your VRAM and RAM and get a list of Ollama models that fit.

Local models are big, and it’s easy to collect dozens of gigabytes without noticing. A few minutes of housekeeping every month or so keeps your disk free, your tools current and your setup predictable.

See what you have

ollama list   # every model, its size, and when it was last downloaded or changed
ollama ps     # what's loaded in memory right now

Ollama keeps its models here, unless you’ve moved them with OLLAMA_MODELS:

SystemModels folder
Windows%HOMEPATH%\.ollama\models
Mac~/.ollama/models
Linux (service install)/usr/share/ollama/.ollama/models

Ollama stores models as shared pieces, so the folder can be smaller than the sizes in ollama list add up to. If two names show the same ID in ollama list, they’re the same model under two names, such as qwen3:latest and qwen3:8b when latest points to the 8B size. Removing one of them frees nothing.

For the rest of your drive, use Settings → System → Storage on Windows (or a free disk-usage tool such as WinDirStat), System Settings → General → Storage on a Mac, or du -sh on Linux.

Remove models you don’t use

ollama rm qwen3:8b

This deletes the model, plus any files that no other model still needs.

  • Good candidates: the same model in several sizes, older generations you’ve replaced, and one-off experiments.
  • Check your apps first. If Continue, Open WebUI or a script still points at a model you’ve removed, it fails with a “model not found” error. Pull it again, or change the app’s setting.

Update models without surprises

Run ollama pull <model> again to update a model. Only the parts that changed are downloaded, and if it’s already current, nothing is.

  • Name the size. A tag like latest can point to a different build after the publisher updates it. qwen3.5:9b is clearer than qwen3.5 (why).
  • Keep a known-good copy of a model you depend on before you update it. The copy keeps the old version’s files until you remove it:
ollama cp qwen3.5:9b qwen3.5:9b-keep
ollama pull qwen3.5:9b
  • Re-test after an update. Run a few of your usual prompts, because replies can change.
  • New generations are new names. Pulling qwen3 again won’t turn it into qwen3.5. Browse the Ollama library now and then, or check the GPU picker for current suggestions.

To update every model at once in PowerShell:

ollama list | Select-Object -Skip 1 | ForEach-Object { ollama pull ($_ -split '\s+')[0] }

Or on a Mac or Linux:

ollama list | tail -n +2 | awk '{print $1}' | xargs -n1 ollama pull

Models you built yourself with ollama create can’t be pulled, so they show an error. That’s harmless.

Update the apps

AppHow it updates
Ollama (Windows, Mac)Downloads updates in the background. Restart Ollama from its tray or menu bar icon to apply one.
Ollama (Linux)Re-run the install script (Linux guide)
Ollama (Docker)docker pull ollama/ollama, then recreate the container. Models kept in a volume, as in the official instructions, stay put.
LM StudioUpdates from inside the app. Its engines (the llama.cpp and MLX runtimes) update separately, so check those too.
Open WebUIPull the new image and recreate the container (steps)
ComfyUIThe desktop app updates itself; the portable build has update scripts in its update folder

Update Ollama before pulling a brand-new model: new model architectures often need a recent release. Updates also fix bugs and security issues, which matters most if other devices can reach your setup.

Move models to a bigger drive

  1. Quit Ollama.
  2. Move the whole models folder to the new drive.
  3. Point OLLAMA_MODELS at the new location: steps for Windows, Mac and Linux. On Linux, the folder must be owned by the ollama user.
  4. Start Ollama and run ollama list to check everything is there.

Use an SSD if you can. Models load much more slowly from a hard drive or USB stick.

Clear out the hidden caches

Ollama isn’t the only thing filling your drive.

WhatWhereHow to tidy it
LM Studio modelsShown, and changeable, in the My Models tabDelete models from that tab
Hugging Face downloads (Python tools, mlx-lm, fine-tuning)~/.cache/huggingface (Windows: %USERPROFILE%\.cache\huggingface)The hf command-line tool can list and delete cached models; run hf cache --help. Set HF_HOME to move the cache.
ComfyUI modelsThe models folder inside ComfyUIDelete checkpoints and LoRAs you don’t use. extra_model_paths.yaml lets several installs share one folder.
DockerDocker’s own storagedocker system df shows what’s using space; docker image prune removes old unused images
Python environmentsEach project’s venv or conda environmentDelete environments you no longer use; pip cache purge clears pip’s download cache

Be careful with Docker volumes. Open WebUI keeps your chats and settings in a Docker volume, so avoid commands that delete volumes unless you’re sure what’s in them.

Uninstalling completely

  • Windows: Settings → Apps → Installed apps → Ollama → Uninstall. Then check for a leftover .ollama folder in your user folder (and wherever OLLAMA_MODELS pointed), and delete it if you want the space back.
  • Mac: quit Ollama, delete it from Applications, and delete ~/.ollama to remove your models.
  • Linux: the official Linux docs list the commands to remove the service, the ollama user and the models folder.

A monthly routine

  1. Update Ollama and your other AI apps.
  2. Run ollama list and remove what you haven’t used.
  3. Re-pull the models you keep, with a -keep copy of any you rely on.
  4. Re-test your everyday prompts.
  5. Check the Hugging Face, ComfyUI and Docker folders.

Something not working after an update? See When it doesn’t work.

Sources: Ollama FAQ, Ollama Linux docs, Hugging Face cache guide, LM Studio docs, Open WebUI docs.

Get the Local AI weekly

New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.

Subscribe free →

Lila Forge Kits

Local-LLM App Starter

A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.

kits.lilaforge.com →
VibeEngine

Best local LLM tools, ranked

Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.

vibeengine.lilaforge.com →