Ollama, LM Studio & Hugging Face Are Eating Your Disk: How to Manage AI Models on Mac

The AI Model Storage Problem

Running local LLMs on Apple Silicon is genuinely useful. But it comes with a storage cost that sneaks up on you. A typical developer's Mac after 6 months of AI experimentation might contain:

• ~/.ollama/models/ — 15–40 GB of pulled Ollama models
• ~/.cache/huggingface/hub/ — 20–60 GB of HuggingFace model shards
• ~/Library/Application Support/LM Studio/models/ — another 10–30 GB
• ~/mlx-models/ — Apple MLX quantized weights
• ~/ComfyUI/models/ — image generation model checkpoints

Total: easily 50–150 GB of model weights, often with significant duplication between hubs.

Cross-Hub Duplication is Common

Ollama's llama3.2:3b and the HuggingFace meta-llama/Llama-3.2-3B-Instruct GGUF Q4_K_M are often byte-for-byte identical. Downloading through different interfaces creates silent duplicates.

Parsing GGUF Headers to Identify Models

A .gguf file begins with a 4-byte magic number (GGUF in ASCII) followed by a structured header containing:

• Model architecture (LLaMA, Mistral, Gemma, Qwen, DeepSeek, Phi, etc.)
• Context length
• Number of layers and heads
• Quantization type (Q4_K_M, Q8_0, F16, etc.)
• Vocabulary size
• Model author and name from metadata

MacPilot parses these headers directly — without loading the full model into memory — to give you meaningful names and sizes rather than opaque blob filenames.

Finding Duplicate Models

MacPilot computes SHA-256 hashes of GGUF files across all detected model directories and surfaces duplicates with their sizes. You can then use APFS clonefile(2) to replace the duplicate with a block-sharing clone, or simply delete it from the hub you use less.

Which Models Are Safe to Delete?

• Safe to delete: Old quantization versions (Q4_0 when you now have Q4_K_M of the same model), experimental pulls you tested once, superseded model versions.
• Keep: Models actively used in workflows, fine-tuned or LoRA-adapted weights (not re-downloadable), models pulled with specific system prompts stored alongside them.

MacPilot's AI Model Registry shows last-accessed timestamps to help you identify unused models.