How to Run Gemma 4 Locally With Ollama, llama.cpp, and vLLM

TL;DR Google Gemma 4 dropped on April 2 under Apache 2.0 and it's genuinely good: the 31B dense model hit #3 on the Arena AI leaderboard, beating models 20x its size. You can run it locally with Ollama in about two minutes, or go the llama.cpp / vLLM route if you want more control. But there are real bugs right now, especially on Apple Silicon and with tool calling. This guide covers all three options, what hardware you actually need, and the workarounds for the issues I've hit so far....