Tunix trains agentic Gemma 4 on TPUs

20 11 2026
Tunix trains agentic Gemma 4 on TPUs

Google released a Tunix repository that runs multi-turn GRPO training for a Plane Strike agent. It uses Gemma 4 E2B on a single Cloud TPU v6e-8 host to improve spatial reasoning and tool calling. The baseline model wins 9.4% of games, rising to 95% after 100 steps.

It is useful if you want to experiment with agentic RL on Google hardware without building custom infrastructure. The code is tied to Tunix and TPUs, so it will not help if you use other clouds or frameworks. You also need a working GCP setup to test it.

more: https://developers.googleblog.com/multi-turn-agentic-reinforcement-learning-with-tunix-and-gemma-4-on-google-tpus/

my other side: ivo.blogspot.com


Actions

Information

Leave a Reply