Qwen releases QwenGyre online reinforcement learning framework for agents

On October 12, 2026, Qwen officially launched QwenGyre, an online reinforcement learning framework designed for ultra-long-horizon agent tasks. The framework addresses major training bottlenecks in complex agent tasks that involve hours of execution, hundreds of interdependent interaction rounds, and millions of accumulated tokens. QwenGyre features an elastic scheduling mechanism that divides GPUs into multiple computing units, allowing newly freed units to immediately join training while ongoing tasks retain their progress and cache, with a core unit finally aggregating results for parameter updates. Additionally, the framework records real context via Token-in-out, organizes trajectories into trees by shared prefixes, expands only selected paths into training samples, and provides feedback for timed-out yet partially completed tasks through Partial Scoring. Official online reinforcement learning experiments on NL2RepoBench showed that QwenGyre delivers notable performance and efficiency gains.