AI homelab user estimates their setup is 60–70% local

One user describes an AI homelab that defaults to local models, estimating it is 60–70% local with a goal of reaching 100%. A homegrown plugin uses Arch-Router to decide when to switch to cloud models. They report Qwen 3.8 27B sometimes reaching 150-plus tokens per second on a 5090 external GPU. Two DGX Spark machines run Deepseek v4 Flash 0731 and handle background batch jobs and longer development builds. Their setup also includes a Mac mini for development and a Pi 5 for monitoring. The user says they started with just the Mac mini, stress that all this hardware isn't necessary, and regret buying the external GPU.