Tencent T1 focuses on long-horizon Agent tasks: over 300 consecutive operations, benchmark score up 20 points.

动察 Beating AI News Flash: Tencent took Qwen3.5-122B-A10B and specifically trained it for terminal Agent, creating T1. It mainly handles complex tasks in the Linux terminal, and can continuously call tools for more than 300 rounds in a single task. Its Terminal-Bench 2.1 score rose from 43.8% for the base model to 64.0%, an increase of 20.2 points.

Of these 20.2 points, the bulk came from reinforcement learning. SFT only raised the score to 49.4%, and the subsequent reinforcement learning added another 14.6 points. The team prepared about 15,000 terminal tasks, each equipped with automated tests. Even if the Agent did not complete the entire task, as long as it completed some of the requirements, it could still receive corresponding rewards.

Long tasks also have another problem. When the Agent actually does the work and later trains on that experience, the same piece of content may be re-tokenized into different tokens, and MoE may also switch to a different set of experts for processing. T1 records both the tokens generated at the time and the selected experts, and directly replays them during training, reducing this training-inference discrepancy by about one third.