Train Gemma 4 Agents With Tunix on TPUs
Google trained Gemma 4 E2B with Tunix agentic reinforcement learning on one TPU host. Its win rate in Plane Strike rose from 9.4% to 95% after 100 steps.
Google trained Gemma 4 E2B with Tunix agentic reinforcement learning on one TPU host. Its win rate in Plane Strike rose from 9.4% to 95% after 100 steps.
Open models of 32B parameters or fewer now score above GPT-4.5 on a combined benchmark index, and they run offline on a laptop with 16 to 32 GB of memory.