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.
Related event: Tencent Hunyuan Releases T1, an RL-Trained Terminal Agent(2 posts)→
More from coding & agent
- Build Your Own Design Harness: A Recipe to Feed Straight Into Your Coding Agent — floguo · 2026-09-11
- MCP OAuth breaks in Codex after token expiry, as devs slam inconsistent host implementations — max__drake · 2026-09-11
- OnePort scans AI-written code for secrets and CVEs before you ship — 100k free credits — Some_Intention9946 · 2026-09-11
- User has GLM 5.3 Flash research game one-shots: 145 calls, 15M tokens in 48 min — yeah_likerage · 2026-09-11
- CopilotKit open-sources OpenBot: self-owned AI coworkers that hit 4.7k GitHub stars — Roger_M_Taylor · 2026-09-11
- Dev Uses OpenAI Astra With Three.js to Build Graphics Demo, Says AI Is Dissolving Rendering-Math Barrier — algo_diver · 2026-09-11