I ran Gemma 4 as a local model in Codex CLI (blog.danielvaughan.com)

A developer recently ran the Gemma 4 model locally via Codex CLI to evaluate its practicality against cloud-based models for coding tasks, focusing on cost, privacy, and resilience. The testing utilized a 26B MoE variant on a MacBook Pro and a 31B Dense variant on a high-powered Dell PC, with both setups adapted for efficient operation. Notably, Gemma 4's performance marked a leap from a previous 6.6% tool-calling success rate to an impressive 86.4%, demonstrating substantial improvements in its ability to perform agentic coding tasks. The results illustrated that while the Mac generated tokens significantly faster, the Dell system produced higher-quality code with fewer errors and less iterative debugging needed. This finding underscores that for practical coding applications, model accuracy outweighs token generation speed. Ultimately, the conclusion reached allows for a hybrid approach in workflows, utilizing local models for sensitive tasks while reserving cloud models for more complex coding challenges. This development signifies a vital shift for the AI/ML community, as it enhances the viability of local models in professional coding environments.