According to CoinMeta, the QwenGyre proposed by Alibaba's Qianwen team is a reinforcement learning framework specifically designed for ultra-long tasks. This framework addresses the issues of low utilization and redundant training data in traditional training systems by dynamically switching the purpose of GPU. The team verified this approach and found that with the same GPU budget, completing 48 steps of reinforcement learning took 75.42 hours using this new framework, whereas the traditional asynchronous method required 134.55 hours, resulting in a reduction of training time by about 44%. After training, the model's pass rate on the test nl2repobench generated in the code repository increased from 52.48% to 58.54%.