Today, we’re introducing Gemini 4 Hydrogen, our daily driver for agentic workflows
Read in full here:
Are Gemma models good for agentic coding?
Gemma is Google’s open-weight family, so it’s a separate line from Hydrogen rather than a smaller version of it. From what I’ve seen, small open models are fine for single-file edits but get shaky on long tool-calling loops, which is exactly where agentic coding lives. The upside is cost: run locally they’re free per token, so it’s worth trying the biggest one your GPU can hold on a real task from your own repo. (I’m with aicost.tools, so the local-vs-API cost question is my daily bread.)