Gemma 4 on Raspberry Pi 5: A Surprisingly Usable Local AI Setup

A hands-on experiment running Google's Gemma 4 E2B (4B parameter, ~4.5GB) model on a Raspberry Pi 5 with 8GB RAM using LM Studio CLI. The setup involves installing LM Studio headless, downloading the model to an SSD, starting an API server, and using socat to expose it over the local network. The model is then connected to the Zed editor via its OpenAI-compatible API endpoint. Performance tests show response times of 5–6 minutes for typical prompts, with all CPU cores maxed out during generation — usable for non-interactive scripts and automation but too slow for real-time chat.