Gemma 4: A Practical Guide to Running Frontier AI on Your Own Hardware

"This article was originally published on my Substack." There’s a quiet assumption baked into the way most of us use AI today: you type a prompt, it leaves your machine, travels to a data center somewhere, gets processed on hardware you don’t own, and the answer comes back. For most of the last three years, “using AI” has meant “renting AI.” Your data leaves. You hope for the best. Gemma 4 is Google DeepMind’s clearest challenge to that model yet. Recently released under an Apache 2.0 license, it’s a family of four open-weight models. They range from a 2-billion-parameter edge model that fits on a phone to a 31-billion-parameter dense model that runs on a single consumer GPU. These aren’t research toys. The 31B variant currently ranks as the #3 open model in the world on the Arena AI text leaderboard, outcompeting models twenty times its size. The 26B model sits at #6. Built on the same research and technology behind Gemini 3, these models handle multi-step reasoning, native function c