Alibaba unveils QwenGyre, a framework for long-horizon reinforcement learning in intelligent agents

On October 12, 2026, Alibaba's Qwen announced the release of QwenGyre, a framework designed for online reinforcement learning in long-horizon intelligent agent tasks. The framework enables efficient training for tasks that span hours, involving hundreds of interactions between the model and environment, with millions of token processing. QwenGyre optimizes GPU usage by dividing them into multiple computing units. When a task completes, the freed units are redirected to training, while ongoing tasks remain unaffected. The framework also retains previous computations, allowing new units to join training immediately. This approach enhances both performance and efficiency. In benchmark tests, QwenGyre achieved a 1.78× end-to-end acceleration compared to Async, with performance close to Colocate. On Qwen3.6-122B, it improved task completion across NL2RepoBench, DeepSWE, and TerminalBench. The framework also supports Partial Scoring, providing feedback on partially completed tasks, ensuring no effort is wasted.