Where Ollama, LM Studio and Hugging Face store models on a Mac

Q: Where does each app keep its models?

Ollama: ~/.ollama/models, split into blobs (the weights, named by their SHA-256) and manifests. The OLLAMA_MODELS environment variable moves it elsewhere. LM Studio: ~/.lmstudio/models, one folder per publisher and model. If you picked another models folder in LM Studio, possibly on an external drive, models are there, and older downloads may still be in the default folder. Hugging Face: ~/.cache/huggingface/hub, one models--publisher--name folder per model. Python libraries, MLX and many apps download here. HF_HOME moves it. Datasets go to ~/.cache/huggingface/datasets. Others: Whisper in ~/.cache/whisper, PyTorch Hub in ~/.cache/torch.

Q: How do I see and remove Ollama models?

ollama list ollama rm llama3.2:3b ollama list shows each model with its size; ollama rm removes one. Don't delete files in blobs by hand: models share layers, and a missing blob breaks the others.

Updated November 1, 2026

Local AI models are some of the biggest files on a developer's Mac, and each app keeps its own copy in its own place.

The short answer: Ollama keeps models in ~/.ollama/models, LM Studio in ~/.lmstudio/models (or the folder chosen in its settings), and Hugging Face in ~/.cache/huggingface/hub. Remove Ollama models with ollama rm, never by deleting its files.

Where does each app keep its models?

Ollama: ~/.ollama/models, split into blobs (the weights, named by their SHA-256) and manifests. The OLLAMA_MODELS environment variable moves it elsewhere.
LM Studio: ~/.lmstudio/models, one folder per publisher and model. If you picked another models folder in LM Studio, possibly on an external drive, models are there, and older downloads may still be in the default folder.
Hugging Face: ~/.cache/huggingface/hub, one models--publisher--name folder per model. Python libraries, MLX and many apps download here. HF_HOME moves it. Datasets go to ~/.cache/huggingface/datasets.
Others: Whisper in ~/.cache/whisper, PyTorch Hub in ~/.cache/torch.

How much space do local models take?

Roughly the parameter count times the bytes per weight. A 7 to 8 billion parameter model quantized to 4 bits is about 4 to 5 GB; a 70B model at 4 bits is about 40 GB. The same model at 8 bits is about twice as big, and unquantized weights in 16 bits about four times. A few experiments add up to 100 GB quickly.

How do I see and remove Ollama models?

ollama list
ollama rm llama3.2:3b

ollama list shows each model with its size; ollama rm removes one. Don't delete files in blobs by hand: models share layers, and a missing blob breaks the others.

How do I remove LM Studio and Hugging Face models?

In LM Studio, open My Models and delete the model there, or move its folder to the Trash while LM Studio is closed. For Hugging Face, the official command line lists the cache and lets you pick what to remove:

pip install -U "huggingface_hub[cli]"
huggingface-cli scan-cache
huggingface-cli delete-cache

Newer versions of the library call the same tool hf. Moving a models--… folder to the Trash works too.

Can the same model be stored twice?

Often. Pulling a model in Ollama and downloading it again in LM Studio, or downloading it once through a Python script, leaves identical weights in two or three places. Files of the exact same size are the first hint; matching SHA-256 checksums (shasum -a 256 <file>) prove it. Ollama's blob names already are that checksum.

Is it safe to delete a local model?

Yes, as long as you can download it again: nothing else on the Mac depends on it, apart from apps set to use it. The cost is the download time. Keep models you fine-tuned, merged or converted yourself: those can't be downloaded again.