# Delete Ollama and LM Studio Models on Mac Safely > Remove LM Studio models and clear Ollama model storage on Mac while protecting chats, private checkpoints, adapters, and shared blobs. Published: 2026-07-18 | Updated: 2026-08-24 To delete an LM Studio model or clear Ollama's model storage, start with each tool's own inventory instead of removing blobs in Finder. Local AI tools can fill a disk quickly because model weights are the product, not a small cache around it. A model may be downloadable again, but the download can be expensive, a revision can disappear, or its license and access can change. Inventory models through the owning tool and preserve chats, fine-tunes, adapters, prompts, and project data separately. ## Where local AI tools store their weight The models are almost always the heavy part, and each one is measured in gigabytes. - **Ollama** keeps models in `~/.ollama/models` by default, or in the directory set by `OLLAMA_MODELS`. Its current [CLI reference](https://docs.ollama.com/cli) uses `ollama ls` to inventory models and `ollama rm ` to remove one by name. This is safer than deleting files by hand because the tool can update its own model store. - **LM Studio** lets you choose the models directory. Use My Models in the app or [`lms ls`](https://lmstudio.ai/docs/cli/local-models/ls) to list local models and sizes, then remove a model through LM Studio so the app's catalog remains consistent. Do not rely on one hardcoded hidden path. - **Hugging Face** libraries cache downloads in `~/.cache/huggingface/hub`, which grows as scripts pull models and datasets unless `HF_HOME` or another cache setting changes the root. `hf cache ls` inventories repositories and revisions. Preview an exact removal with `hf cache rm model/ --dry-run`, or preview detached and incomplete revisions with `hf cache prune --dry-run`. The [Hugging Face cache guide](https://huggingface.co/docs/huggingface_hub/en/guides/manage-cache) documents both commands. Check the totals directly so you know what you are dealing with: ``` du -sh ~/.ollama/models ~/.cache/huggingface 2>/dev/null ``` ## Cache and data are not the same thing This distinction keeps cleanup safe. A published **model** may be downloadable again, but a local fine-tune, adapter, converted quantization, or private checkpoint may not be. Desktop AI apps can also keep conversation state and attachments in Application Support or synchronize some history to an account. Check the app's export and sync behavior rather than assuming either that every chat is local or that every local file is disposable. Model caches are large, but they are not free to recreate. ## Removing an AI app completely If you are done with a tool, the app bundle is the small part. To reclaim the real space you have to remove its models and support files too, the same problem as [uninstalling any app completely](https://mole.fit/blog/how-to-completely-uninstall-apps-on-mac). For Ollama, remove named models with `ollama rm` before uninstalling the app. For LM Studio, remove models from its configured directory through the app. For a desktop chat client, export needed history and follow the vendor's uninstall guidance before reviewing support data. Do not delete an entire Application Support container merely because its name matches an app. ## Under the hood: how the tools store models, and why size is inherent Ollama keeps each model as content-addressed blobs (files named by their sha256 hash) under `~/.ollama/models/blobs`, with small manifests that link a model name and tag to the blobs it needs. Two models built on the same base layer share blobs, so `ollama rm` deletes a manifest first and frees a blob only once nothing else references it. That is exactly why deleting blob files by hand is risky: you can orphan a model that still points at them. The size is not overhead, it is the model: billions of parameters still require gigabytes even after quantization, and larger or higher-precision models require far more, because the weights are the thing. Hugging Face's cache works the same way, storing immutable snapshots keyed by revision so repeated pulls deduplicate. This is why the rule is "remove through the tool's own command": it keeps the content-addressed store consistent instead of leaving dangling blobs.
Two named model manifests point into a shared pool of hash-named blobs, one blob referenced by both, so removing a model frees a blob only when it is no longer referenced.
Local model tools use a content-addressed store: models share blobs, so removing one frees space only for the blobs nothing else references, which is why the tool's own remove command keeps it consistent.
## Where a disk map helps A disk map such as [Mole](https://mole.fit/)'s Analyze view can reveal an unexpected model store, but model names, shared blobs, and revision references belong to Ollama, LM Studio, or Hugging Face. A general cleaner should avoid model stores and chat transcripts unless the user identifies an exact target. ## A safe order of operations List models with their owning tool, distinguish published weights from private or locally transformed artifacts, export unique conversations and adapters, then remove one named model at a time. Recheck disk use because shared blobs may limit the space reclaimed. Remove whole support directories only as part of a verified uninstall with no data left to preserve. --- Canonical HTML page: https://mole.fit/blog/how-to-remove-ai-tool-leftovers-mac Blog index for agents: https://mole.fit/blog/llms.txt Site index for agents: https://mole.fit/llms.txt