The AI Engineer's Model Disk Cleanup Guide: Hugging Face, Ollama & PyTorch

Bottom Line: Artificial intelligence developers, machine learning researchers, and local LLM hobbyists using Ollama, LM Studio, Hugging Face Transformers, PyTorch, and ComfyUI often find their SSDs filled with multi-gigabyte GGUF, Safetensors, and PyTorch model weights. Running this cleanup guide reclaims 30 GB to 200 GB+ of storage space safely.

The AI Model Weight Storage Crisis

Local AI models (Llama 3, Mistral, DeepSeek, Stable Diffusion XL, Whisper) are massive. A single 70B quantized GGUF model weighs 40 GB, while Stable Diffusion checkpoint weights (.safetensors) weigh 7 GB each.

Because model loaders silently save models to hidden home directories (~/.cache or ~/.ollama), developers frequently run out of SSD space without knowing where their storage went.

+-------------------------------------------------------------------------+
| [System: 30GB] [Ollama GGUF & Hugging Face Weights: 180GB] [Free: 2GB]  |
+-------------------------------------------------------------------------+
                                        ^
                              Hidden AI Model Bloat!

Tool-by-Tool AI Storage Audit & Cleanup Commands

1. Ollama GGUF Model Store (~/.ollama/models) — 20 GB to 100 GB

Ollama saves pulled models (ollama run llama3) as raw blobs inside ~/.ollama/models/blobs.

How to List & Remove Ollama Models:

# List all locally installed Ollama models
ollama list

# Remove unused models
ollama rm llama3:70b
ollama rm mistral:7b

# Purge leftover blob manifests
rm -rf ~/.ollama/models/manifests/*

2. Hugging Face Hub Cache (~/.cache/huggingface/hub) — 20 GB to 80 GB

The transformers and diffusers Python libraries store downloaded model snapshots in ~/.cache/huggingface/hub.

How to Purge via CLI:

# Install Hugging Face official cache management CLI
pip install huggingface_hub

# Delete cached models interactively
huggingface-cli delete-cache

# Or force purge all cached models
rm -rf ~/.cache/huggingface/hub/*

3. PyTorch Hub Checkpoints (~/.cache/torch/hub) — 10 GB to 40 GB

# Clear downloaded PyTorch vision and audio model weights
rm -rf ~/.cache/torch/hub/checkpoints/*

4. vLLM & LM Studio Model Caches — 15 GB to 60 GB

# Clear vLLM model cache
rm -rf ~/.cache/vllm/*

# Clear LM Studio downloaded model folder
rm -rf ~/.cache/lm-studio/models/*

5. ComfyUI & Automatic1111 Stable Diffusion Checkpoints — 20 GB to 100 GB

Check your ComfyUI/models/checkpoints or stable-diffusion-webui/models/Stable-diffusion folders for duplicate .safetensors files.

AI Storage Safety & Impact Matrix

Frequently Asked Questions (FAQ)

Will deleting Hugging Face cache break my Python scripts?

No. If a script calls AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b"), Hugging Face will simply re-download the model weights if they are missing from local cache.

Related AI Guides

• Learn about vLLM Model Cache.
• Learn about Whisper Model Cache.