I ran Gemma 4 as a local model in Codex CLI

2026-04-15

An evaluation of Gemma 4 as a local agentic coding model in Codex CLI, compared against cloud models like GPT-5.4, showed that first-pass reliability, not raw token speed, was the most important performance metric, as cloud models were the fastest and most reliable. While cloud models still offer the best overall performance, Gemma 4's tool-calling ability makes local agentic coding viable.