Local AI Models Eating Your Disk: Ollama, LM Studio (2026)
Local model files are among the biggest reclaimable items on a 2026 developer Mac. Ollama stores models in ~/.ollama/models, LM Studio in ~/.lmstudio/models, and anything pulled from Hugging Face lands in ~/.cache/huggingface/hub. Each model is multiple gigabytes and every one is re-downloadable, so deleting the ones you no longer use is completely safe.
Why local models got so heavy
Running models on your own Mac exploded in popularity once Apple Silicon made it practical. The catch: every model you try is a multi-gigabyte GGUF or safetensors file, and the tools keep them forever by default. Pull a dozen models to “compare” them and you're suddenly down 60–100 GB you forgot about.
Where each tool hides its models
| Tool | Location | Reclaim with |
|---|---|---|
| Ollama | ~/.ollama/models | ollama rm <model> |
| LM Studio | ~/.lmstudio/models | Delete inside the app, or remove the folder |
| Hugging Face cache | ~/.cache/huggingface/hub | huggingface-cli delete-cache |
| llama.cpp / manual GGUF | wherever you saved them | plain rm |
All of these are download caches. Delete a model and the tool simply fetches it again next time you ask for it.
Cleaning up Ollama
List what you've got and its sizes, then remove what you don't use:
ollama list
ollama rm llama3.1:70b
The 70B-class models are the space eaters — a single one can be 40 GB. If you only reach for a small model day-to-day, the big ones are pure dead weight until you next need them.
The Hugging Face cache is the sneaky one
Anything you ran through transformers, diffusers or a from_pretrained call was cached under ~/.cache/huggingface — including models you only used once. It's often the single largest hidden folder on an ML machine.
huggingface-cli scan-cache shows what's there; huggingface-cli delete-cache gives you a checklist to prune it. Everything removed is re-downloadable.
Finding it without hunting
MacCleaner surfaces these model caches as their own group with real measured sizes, so you can see that ~/.ollama is holding 80 GB before you decide. It's read-only until you act, keeps deletions recoverable, and never pre-selects anything. Pair it with the node_modules guide and Xcode guide for a full developer sweep.
Frequently asked questions
Is it safe to delete Ollama models?
Yes. Use ollama rm to remove them, or delete ~/.ollama/models. Ollama re-downloads any model on demand, so you only lose the time to fetch it again.
How much disk can local AI models use?
On an active 2026 ML/dev Mac, 50–150 GB across Ollama, LM Studio and the Hugging Face cache is common — a single large model can be 40 GB by itself.
Will clearing the Hugging Face cache break my projects?
No, as long as you have an internet connection for the next run. Cached models are re-fetched automatically; only locally-trained weights you never saved elsewhere are unique.
See where your space actually went
MacCleaner scans read-only, shows you every file before anything is touched, and lets you choose Trash or permanent. Free for up to 500 MB of cleanup a day.
Download for Mac