Alibaba Qwen Launches QwenGyre, an Online RL Framework for Long-Horizon Agent Tasks

On Oct. 12, 2026, Alibaba Qwen officially released QwenGyre, an online reinforcement learning framework designed for long-horizon agent tasks. The framework targets training scenarios in which an agent executes complex tasks over several hours, involving hundreds of interdependent interactions with the environment and millions of tokens in total.

According to the company, QwenGyre uses elastic scheduling and trajectory processing to divide GPUs into computing units. As a batch of tasks finishes, available units can immediately switch to training, while running tasks keep their state and progress, and cached computations can be reused without restarting. Newly freed units join the ongoing training round, and a central unit aggregates results to perform a parameter update.

For sample organization, QwenGyre records the real context of each output segment using a Token-in Token-out approach, organizes trajectories into a tree by shared prefixes, and expands only the selected paths as training samples. Partially completed tasks that time out can still provide useful feedback through Partial Scoring.

In the NL2RepoBench online reinforcement learning experiment, with roughly 700,000 tokens per rollout, the model's pass rate increased from 52.5% to 58.5% after 48 training steps. With the same GPU budget, QwenGyre achieved 1.78x end-to-end speedup over an Async baseline, while its training scores were close to those of Async and Colocate. In cross-task validation on Qwen3.6-122B, end-to-end speedups were observed across NL2RepoBench, DeepSWE, and TerminalBench.