Running Gemma 4 Locally with the Codex CLI: What Works (codex.danielvaughan.com)

Gemma 4, Google’s latest open-weights model, is now operational locally with the Codex CLI, marking a significant breakthrough in AI tooling. Previously, earlier iterations like Gemma 3 struggled with functionality, but Gemma 4 boasts a remarkable score of 86.4% on the tau2-bench function-calling benchmark, up from just 6.6% for its predecessor. This leap is attributed to six new special tokens designed specifically for structured function calling, enabling better tool interaction and execution through the Codex CLI’s fixed toolbox. This guide provides comprehensive instructions for getting Gemma 4 running on two setups: a MacBook Pro with 24 GB of unified memory and a Dell GB10 with 120 GB. Each tested setup confirms that Gemma 4 achieves successful tool calling and text generation, a notable advancement for developers focusing on local AI deployment. The reported performance metrics illustrate the efficiency and capability enhancements compared to previous models, reinforcing Gemma 4's potential to reshape AI applications in coding and automation. The guide also highlights pitfalls and inefficiencies encountered during testing, ensuring users avoid common errors and fully leverage the capabilities of this new model.