Reasoning traces help make language models more capable, trustworthy, transparent, and valuable. But building models that reason well requires more than just good data. It takes open tools, strong foundations, and training approaches that the open-source world doesn’t yet have in one place.

In this hackathon, you’ll start with Google’s open-weight Gemma model (Gemma2 2B or Gemma3 1B), fine-tune it with Tunix on TPU, and teach it how to reason through complex questions. You’ll create a model that not only gets the right answer, but also explains how it got there.

Your solution should include a working training pipeline using Tunix and Gemma. Share your configs, reward function composition, and recipes so others can reproduce your results and build on them.

This contribution can help make step-by-step reasoning more accessible for the entire open-source community and lower the barrier to building transparent, capable, and explainable AI.

Like our parents always said: think first, talk later. Let’s teach our models to do the same.