Running Google’s new Gemma 4 model on Anthropic’s Claude Code delivers open-source, high-end coding power without the hardware hassle. The 31B variant hits 80 percent on LiveCodeBench v6, 89.2 percent on AIME 2026 (up from Gemma 3’s 20.8 percent), and a 2150 Codeforces ELO — levels you’d expect from closed models just months ago. It also nails 85.2 percent on MMLU Pro, showing strong reasoning beyond pure syntax. Both the 31B and 26B sizes handle up to 256K tokens, so you can load entire repositories in one prompt. Gemma 4 can invoke external tools and APIs directly, smoothing out multi-step workflows in Claude Code. Unlike its predecessors, Gemma 4 ships under Apache 2.0, so you get unrestricted commercial use, fine-tuning freedom, and easy deployment. You can run it locally through Ollama or tap into Blackwell GPUs via Ollama Cloud. Integrating it with Claude Code’s agent framework means the model can call scripts, execute tests, and chain operations without extra glue code. Getting started takes five steps. First, install Ollama (brew on Mac, shell on Linux, or Windows installer) and Claude Code v2.1+ (curl script on Unix, PowerShell on Windows). Then pull the cloud model with ollama pull gemma4:31b-cloud and launch Claude Code against it via ollama launch claude --model gemma4:31b-cloud. A quick /status check shows “Model: gemma4:31b-cloud” and your local Anthropic endpoint. From there, you’re set to feed long codebases straight into Claude Code and tap Gemma 4’s new capabilities.