Researchers have introduced T1, a 122B parameter Mixture-of-Experts (MoE) model specifically designed for long-horizon terminal tasks such as coding. The model operates within a cloud-based sandbox shell, capable of executing over 300 tool-call turns per task, with success measured by task-specific verifiers. The training methodology utilizes a specialized reinforcement learning recipe to stabilize actor-critic learning. This includes an aggressive warm-start phase using dense process rewards based on the number of passing verifiers. To address training instability, the authors implemented Two-step Iterative Token Optimization (TITO) and Rollout Routing Replay (R3). TITO utilizes drift repair at turn boundaries to train on sampled token identifiers, while R3 replays per-token expert choices from the MoE layers during training. These techniques successfully reduced the training-to-inference log-probability difference from 0.021 to 0.013. To prevent overfitting, the model was trained on an out-of-distribution corpus comprising isolated seeds and synthesized tasks distinct from Terminal-Bench 2.1. Experimental results demonstrate significant performance gains: T1 improved the base model's performance on Terminal-Bench 2.1 from 43.8% to 64.0%. Furthermore, on the Long-Horizon Terminal Bench, T1 achieved a 27.9% resolution rate, outperforming larger models such as GPT-5.4 and GLM-5.1.