huggingface.co web signal

T1: 122B MoE Terminal Agent Reaches 64% on Terminal-Bench 2.1

TL;DR

  • T1 is a 122B Mixture-of-Experts model trained with reinforcement learning to drive a real shell in a cloud sandbox for 300+ tool-call turns per task.
  • The post-train pipeline lifts Terminal-Bench 2.1 from a 43.8% base to 64.0% resolved, and reaches 27.9% on Long-Horizon Terminal Bench.
  • TITO construction and rollout routing replay cut the training-to-inference log-probability difference from 0.021 to 0.013 on an MoE actor.

T1, a 122B Mixture-of-Experts model, resolves 64.0% of Terminal-Bench 2.1 tasks after reinforcement-learning post-training that lifted a 43.8% base. That is the headline claim in the T1 paper posted to Hugging Face by Junyao Yang and six collaborators.

The setup is blunt. The model is 'operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier,' the authors write. No reward model in the loop; a task passes if its verifier says it passes.

Two named techniques do the work on the training-inference gap that has haunted on-policy MoE RL. TITO trains on 'the exact sampled token identifiers with drift repair at turn boundaries.' Rollout routing replay records 'the sampler's per-token expert choices at every MoE layer' and replays them during training. Together, the paper reports, they 'cut the training-to-inference log-probability difference from 0.021 to 0.013.'

On the harder Long-Horizon Terminal Bench, T1 reaches 27.9% and, per the authors, 'surpasses GPT-5.4 and GLM-5.1.' The abstract does not report cost figures, latency, or whether weights will be released.

Terminal-agent work is stacking up on our tracker: OpenAI's Agents API and the DeepSeek harness sandbox bug both landed the same week.