Ollama: update, delete a model, uninstall

#Why maintain Ollama

Ollama evolves along with the llama.cpp engine it bundles: each update adds support for new models, speed improvements, GPU fixes, and sometimes format changes. Staying on an old version means missing recent models (« this model is not supported by your version of Ollama ») and carrying bugs that have already been fixed.

The other half of maintenance is disk space. A 7B model in Q4_K_M weighs about 4–5 GB, a 14B model around 9 GB, and a 32B model nearly 20 GB. Test about ten of them over a few weeks and you can quietly fill 100 GB without realizing it. Knowing how to delete them cleanly and check the space actually used is part of the job.

#Check your version

First, check where you stand. The command works identically on all three systems and returns the installed version number.

ollama --version
# ollama version is 0.x.y

Compare this number with the latest version published on the official releases page (github.com/ollama/ollama/releases). If you are several versions behind, the update below takes two minutes.

#Update Ollama (Windows, macOS, Linux)

The update never touches your models: the ~/.ollama directory is independent of the binary. You can therefore update the application without downloading anything again. The method simply depends on your operating system.

1. 01
   Windows

   The application checks for updates automatically and displays a notification when a new version is ready; just click to install it. To force an update, download the OllamaSetup.exe installer again from ollama.com/download and run it over the existing installation: your models are preserved.
2. 02
   macOS

   Same idea: the app offers the update in the menu bar (the Ollama icon → “Restart to update”). Otherwise, redownload the .dmg from ollama.com/download and replace the app in /Applications. If you installed it via Homebrew, use “brew upgrade ollama” instead.
3. 03
   Linux

   Simply rerun the official installation script: it detects the existing installation and replaces the binary with the latest version, without deleting your models or your systemd service.

# Réexécuter le script officiel met à jour le binaire en place
curl -fsSL https://ollama.com/install.sh | sh

# Vérifier, puis redémarrer le service si besoin
ollama --version
sudo systemctl restart ollama

#Delete a model: ollama rm

To clean things up, start by listing what you actually installed. The SIZE column shows the apparent size of each model, and TAG gives the exact name to pass to the removal command.

# Voir tous les modèles installés, avec leur taille
ollama list

# Supprimer un modèle précis (nom:tag exact)
ollama rm llama3.1:8b

# Supprimer plusieurs modèles d'un coup
ollama rm gemma3:12b qwen2.5:7b

Watch the tag: “llama3.1” and “llama3.1:8b” may refer to different entries. Copy the exact name displayed by ollama list so you do not delete anything other than what you intend to target. A model being used by a request will be released when that request ends.

#The reality of shared blobs

Ollama does not store one monolithic file per model. Internally, each model is described by a manifest that references several blobs: the weights (the large GGUF file), as well as the prompt template, parameters, and license. These blobs are identified by their SHA-256 fingerprint and deduplicated: two models that share exactly the same weight file store it only once.

Specifically, if you pulled “llama3.1:8b” and “llama3.1:8b-instruct-q4_K_M,” which point to the same weight blob, deleting either one frees up nothing on the weight side: the other still references the blob. Conversely, deleting the last model that referenced a blob finally removes it from disk. That’s why the space recovered by a ollama rm est can sometimes be much smaller than the size reported by ollama list.

Manifest

A small JSON file for each model:tag, in ~/.ollama/models/manifests/. This is what ollama rm supprime reads first.

Blob

A data file (weights, template, params) named sha256-<fingerprint>, in ~/.ollama/models/blobs/. Shared by all models that reference it.

Deduplication

A blob is deleted only when no manifest references it anymore. That’s a good thing: it prevents duplicate gigabytes of identical data.

#Find and clean up disk space

To see what is really using your disk, check the size of the storage directory instead of relying only on the output of ollama list—the latter adds up sizes that may share blobs.

# Taille totale réellement occupée sur le disque
du -sh ~/.ollama/models

# Détail : blobs (les poids) vs manifests (les métadonnées)
du -sh ~/.ollama/models/blobs ~/.ollama/models/manifests

# Les plus gros blobs en premier
ls -lhS ~/.ollama/models/blobs/ | head

# Taille totale du dossier de modèles
Get-ChildItem "$env:USERPROFILE\.ollama\models" -Recurse |
  Measure-Object -Property Length -Sum |
  Select-Object @{n='Go';e={[math]::Round($_.Sum/1GB,2)}}

The most effective cleanup routine: list the models, identify the ones you no longer use, delete them with ollama rm, then measure again with du -sh to see the recovered space. Because deduplication is involved, delete redundant variants of the same model last—that is where you free up the most space.

#Completely uninstall Ollama

Uninstalling takes two steps: remove the application, then delete the remnants (models and configuration) if you really want to clean everything up. The second step is the same regardless of the OS — you need to delete the ~/.ollama directory.

1. 01
   Windows

   Go to Settings → Apps → Installed apps, search for “Ollama,” and click Uninstall. Then delete the C:\Users\<you>\.ollama folder to remove the downloaded models.
2. 02
   macOS

   Drag Ollama.app from /Applications to the Trash. Then delete ~/.ollama for the models. If you installed it via Homebrew: “brew uninstall ollama”.
3. 03
   Linux

   Stop and disable the service, remove the binary and the systemd file, delete the system user “ollama” created by the script, then delete the models folder.

# 1. Arrêter et désactiver le service
sudo systemctl stop ollama
sudo systemctl disable ollama
sudo rm /etc/systemd/system/ollama.service

# 2. Retirer le binaire (souvent dans /usr/local/bin ou /usr/bin)
sudo rm $(which ollama)

# 3. Supprimer l'utilisateur et le groupe système créés par le script
sudo userdel ollama
sudo groupdel ollama

# 4. Effacer les modèles et la configuration
rm -rf ~/.ollama
sudo rm -rf /usr/share/ollama/.ollama

#Troubleshooting

“this model is not supported by your version”

Your Ollama is too old for the model format. Update it (section above), then restart ollama pull.

ollama rm ne libère presque rien

The weight blob is still referenced by another model variant. Run ollama list and also delete the redundant variants.

“Error: model is currently loaded”

A request is still using the model. Wait for it to finish, or restart the daemon (systemctl restart ollama / quit and relaunch the app) before deleting.

Disk still full after deletion

Check the models' actual location (the OLLAMA_MODELS variable, or /usr/share/ollama/.ollama on Linux when running as a service) and measure again with du -sh.

The ollama command is missing after a partial uninstall

The binary was removed, but the PATH still references it; open a new terminal, or check that no old link remains in /usr/local/bin.

#Go further

The maintenance guide addresses “how to keep Ollama clean”; these guides cover before and after—(re)installing according to your OS, understanding the basic commands, and choosing models worth the space they take.