9B RL Terminal Agent Hits 27% on TerminalBench 2.0, Beats Larger Models

Tmax achieves 27% on Terminal-Bench 2.0 using a 9B parameter model trained with RL, outperforming larger models from prior work. The recipe uses a novel data generation taxonomy combining difficulty control, personas, and verifier diversification to cheaply produce terminal environment training data for both RL and SFT. Code and data are open-sourced.