Open RL Recipe TMax for Terminal Agents Released

TMax, a new reinforcement learning (RL) recipe designed to push terminal agent performance up to frontier-model levels, has been released.

The key highlights are as follows:

• TMax-15k Dataset: A dataset of 14,600 RL environments built through a compositional pipeline that allows precise control over difficulty and diversity. This is more than 2.5x larger than the largest existing open terminal dataset.
• RL Recipe: A simple Outcome-only RL approach is proposed, combining GRPO with stability-improving techniques.

Performance Metrics:

• TMax-9B: Scored 27.2% on Terminal Bench 2.0, achieving top-tier performance among models under 10B. This is close to Claude Haiku 4.5's score of 29.8%.
• TMax-27B: Performance improved to 42.7%, showing results close to those of massive models like Kimi K2.5 (43.2%).