A low-power home server plan for small local AI models

The goal is a low-power, low-cost that can run small AI at home while sending harder work to outside AI services such as Gemini through an API. OpenClaw would act as a chat-based coordinator, with separate profiles and memory for two people on the same machine. A small local model, around 3B to 8B in size, would run through Ollama for quick jobs like summarizing and sorting email.

A larger local model, up to about 14B, is also under consideration for more capable reasoning without using an outside service. The same box would also run , a light media server, a VPN, and Docker hosting for small projects and automations. The media server only needs to match or beat a MacBook Air M2 running Plex on the home network.

A few Gmail accounts would be connected with . The uncertain parts are hardware choice, model size, and compression settings for running the models well.

Key points

  • The target is a low-power, low-cost home server.
  • Small would handle quick email summary and sorting tasks.
  • Harder AI work would be sent to outside services such as Gemini through an API.
  • The same machine would also run , a media server, a VPN, Docker projects, and automations.
  • Media performance only needs to be as good as or better than Plex on a MacBook Air M2 over the .
Read original