Local AI models are quietly filling your SSD
The shift toward running AI locally — for privacy, for latency, for offline use — has quietly introduced one of the fastest-growing categories of disk bloat on developer Macs: model weights.
A single local LLM can range from a few gigabytes to well over a hundred. Multiply that across a few models you downloaded to compare, plus the tools that manage them, and it's easy to lose 50–200GB without ever noticing a single large download that felt significant on its own.
Where it hides
This isn't limited to one app. It's spread across a growing list of tools people install and forget about:
- Model runners and managers — Ollama, LM Studio, and the HuggingFace cache directory.
- Image generation — Stable Diffusion and ComfyUI checkpoints, often several gigabytes each.
- AI coding tools — Cursor, GitHub Copilot, Claude, ChatGPT desktop, Codeium, Tabnine, Continue, Sourcegraph Cody, Windsurf, and Aider all maintain their own local caches.
- ML frameworks — PyTorch, TensorFlow, Conda environments, and Jupyter all accumulate their own cached artifacts over time.
Why generic tools miss this
Most disk-cleanup tools were built before this was a category at all — they know about browser caches and temp files, not about where LM Studio keeps its downloaded checkpoints. Reclaim's AI Cache & Logs view exists specifically for this: it detects and groups cache and model data by tool, so instead of a vague “Other” bucket, you see exactly which AI tool is holding how much, and can clear the ones you're not actively using.
See exactly what’s using your disk space.