Tunix trains multi-turn agents with Gemma 4 on TPUs
This technical guide demonstrates how to train a multi-turn, tool-using LLM agent using Tunix's agentic reinforcement learning pipeline on Google Cloud TPUs. Using Gemma 4 E2B on the Plane Strike board game, the authors show that after 100 steps of agentic GRPO, the model's win rate increases from 9.4% to 95% while reducing average game length. The pipeline utilizes modular components including customized environment wrappers, chat template parsers, and asynchronous vLLM rollout engines. Open article →