How to Delete Abandoned Ollama & Hugging Face Models on Mac

Unlike standard applications, AI frameworks do not use macOS Application Support standard paths with clear file names. Instead, weights are split across hashed blob stores:

First, list all installed Ollama models and their disk footprint:

ollama list

Remove any models you experimented with and no longer use:

ollama rm llama3.1:70b ollama rm mistral-nemo ollama rm deepseek-coder:6.7b

Tip: If you uninstalled Ollama previously without deleting models, the weights might still linger in ~/.ollama/models/blobs. Check with: du -sh ~/.ollama.

Python developers using transformers or diffusers accumulate gigabytes in Hugging Face's snapshot storage. The official CLI includes a cache manager:

huggingface-cli delete-cache

This launches an interactive TUI showing each cached repo, its size, and last modified date. You can select revisions to delete safely without corrupting shared symlinks.

Manual clean (Nuclear option):

rm -rf ~/.cache/huggingface/hub/*

Pretrained weights downloaded by torchvision or timm reside in:

rm -rf ~/.cache/torch/hub/checkpoints/*

For ComfyUI or Automatic1111 users, check your checkpoints subfolder:

find ~/ComfyUI/models/checkpoints -name "*.safetensors" -size +1G

Hunting through 5 different CLI tools and cryptic SHA-256 hash files is tedious and risky. SpaceWise introduces the Model Graveyard:

• Unified AI Scan: Aggregates Ollama, Hugging Face, LM Studio, and PyTorch in one visual screen.
• Human-Readable Names: Decodes SHA hashes back to their actual model names (e.g. "Qwen-2.5-Coder-32B").
• Staleness Detection: Flags weights that haven't been loaded in 60+ days.
• Trash-Only Deletion: Moves weights to macOS Trash so you never suffer accidental data loss.