**[Alibaba's QwenGyre Reduces 2.4T Model Reinforcement Learning Time by 44%]** Golden Finance reports that Alibaba's Qwen team has introduced QwenGyre, a reinforcement learning framework specifically designed for ultra-long tasks in AI agents. As agents begin to undertake complete software engineering tasks, a single execution may last for hours and involve hundreds of model calls and tool operations. Models need to learn from the actual results of these tasks, but traditional training systems often face prolonged waiting times and redundant computations. The first issue is GPU utilization. Traditional synchronous training requires all agents in a batch to complete their tasks before the model can be updated collectively. Since task durations vary significantly, tasks that finish earlier leave some GPUs idle for extended periods. While asynchronous training separates task execution from model training, it fixes GPU allocations for both, making dynamic adjustments difficult. QwenGyre allows GPUs to dynamically switch purposes based on task progress. As tasks are completed, the system reallocates some GPUs to model training while the remaining GPUs continue supporting unfinished agents. This process does not interrupt ongoing tasks, and workspace and tool states are preserved. Newly added GPUs can directly participate in the current batch of training. Another issue is redundant training data. Agents repeatedly retry, compress historical records, or call sub-agents, leaving behind overlapping execution paths. QwenGyre records the actual inputs and outputs of each model call, organizing different paths into a trajectory tree with shared histories. During training, only a limited number of paths are selected, and the same segment of model output is not recalculated. Even if tasks time out, completed and verifiable portions of results can still be used for training. The team validated QwenGyre on Qwen3.8-2.4T, processing approximately 700,000 tokens per task. Under the same GPU budget, QwenGyre completed 48 steps of reinforcement learning in 75.42 hours, compared to 134.55 hours with traditional asynchronous methods—a reduction of about 44% in training time. After training, the model's pass rate on the NL2RepoBench test for code repository generation improved from 52.48% to 58.54%. The paper also shows that QwenGyre achieves training scores comparable to traditional training methods.