T1

A 122B-parameter MoE terminal agent from Tencent's Hy Foundation Model Frontier team (with the National University of Singapore, Georgia, Indiana, and Maryland), obtained by post-training Alibaba's Qwen3.5-122B-A10B purely with reinforcement learning on executed outcomes: the agent operates a real shell in a cloud sandbox for 300-plus tool-call turns per task and is rewarded by each task's own verifier. The recipe is the contribution: an aggressive warm start with a dense process reward that scores trajectories by the number of passing verifiers; a stabilisation stack for sparse agentic RL that trains on the exact sampled token identifiers with drift repair at turn boundaries and replays the sampler's per-token expert routing during training; and a fully out-of-distribution corpus of isolated seeds and synthesised tasks disjoint from Terminal-Bench 2.1.

On Terminal-Bench 2.1 under a fixed harness the pipeline lifts the base model from 43.8% to 64.0%, which the paper places above GPT-5.4 (54.8%), DeepSeek-V4-Flash (56.9%), and Claude Opus 4.6 (63.8%) and close to Claude Opus 4.7 (66.1%), the best result in its size band. The paper does not release weights as of mid-September 2026.