Tmax

Ai2's terminal-agent model family — "a simple recipe for terminal agents" billed as the strongest open RL recipe for terminal agents to date. Sizes 2B/4B/9B/27B fine-tuned from Qwen 3.5 (2–9B) and Qwen 3.6 27B with vision heads removed, trained with DPPO (Divergence Proximal Policy Optimization). tmax-9b scores 27.2 and tmax-27b 42.7 on Terminal-Bench 2.0. Radically open: Apache-2.0 weights plus the full SFT data (TMax-15K, tmax-sft) and all RL rollouts and logprobs published — per-stage ablation models included.