Google's Gemma 4B model passes 2 of 3 Linux sysadmin tasks in QEMU sandbox test

A developer built an open-source evaluation harness called local-agent-sandbox to test whether Google's 4-billion-parameter Gemma model could autonomously diagnose and fix Linux server failures. The setup used QEMU virtual machines with QCOW2 copy-on-write overlays instead of Docker containers, providing stronger isolation and allowing each test environment to be reset in under a second. Gemma was given root SSH access to a headless Ubuntu 24.04 VM and tasked with resolving three distinct DevOps failure scenarios involving Nginx. The model succeeded in two cases — clearing a port conflict and fixing a configuration syntax error — but failed the third by misidentifying a file permission issue and repeatedly rewriting virtual host configs rather than correcting the chmod setting. The experiment concluded that small quantized models can handle real CLI-based system tasks, but require guardrails to prevent them from doubling down on incorrect diagnoses.