T1: RL-trained 122B MoE terminal agent lifts Terminal-Bench 2.1 from 43.8% to 64.0%

heghbalz · x · 2026-09-11

Researchers introduce T1, a 122B-total/10B-active MoE model trained with reinforcement learning to operate a real shell in a cloud sandbox for 300+ tool-call turns per task.

Key results

Terminal-Bench 2.1: 43.8% base → 49.4% SFT → 64.0% T1, a 46.1% relative gain from post-training
27.9% on Long-Horizon Terminal-Bench, surpassing GLM-5.1

Training recipe

PPO on 15K audited tasks with per-assertion verification, rewarding partial progress inside failed trajectories
Aggressively warm-started critic for credit assignment; TITO construction and rollout routing replay preserve sampled tokens and expert choices, cutting the train-to-inference log-prob gap from 0.021 to 0.013
Fully out-of-distribution corpus disjoint from Terminal-Bench 2.1, showing genuine capability transfer

Paper, project page, and data repo are public. The authors note this echoes DeepSeek-V4.1's finding that better data and environment pipelines unlock large gains with established RL methods.

Original post →

T1: RL-trained 122B MoE terminal agent lifts Terminal-Bench 2.1 from 43.8% to 64.0%

More from coding & agent