Qwen’s QwenGyre framework speeds up training AI agents on hours-long tasks

Alibaba’s Qwen team has published QwenGyre, a framework for training agents with reinforcement learning, in which a model improves by being rewarded for good results. It targets what the paper calls extreme-long-horizon tasks: jobs in which one run may take hours, with hundreds of exchanges between model and environment and close to 1M tokens. The paper, on arXiv, has 12 authors from Alibaba Token Hub at Alibaba Group, the University of Science and Technology of China and Tsinghua University.

QwenGyre has two core parts. Its Elastic Scheduler shifts GPUs back and forth between training and rollouts as the load changes, without stopping runs already in progress. Rollouts are the runs in which the agent attempts tasks to produce training data. Its Trajectory Processor pieces together agent histories that branch, gives credit for partial progress and strips out repeated paths so training costs stay in check. The framework works with existing agent tools without changing them, and every experiment in the paper used Claude Code 2.1.220.

On Qwen 3.8 2.4T, Qwen’s flagship model, with 700k tokens per rollout, the same number of training steps took 75.42 hours with QwenGyre, against 134.55 hours for an asynchronous baseline and 91.47 hours for a shared-GPU baseline.

Across all the paper’s setups, QwenGyre ran 1.38 to 1.78 times faster than the asynchronous baseline and 1.21 to 1.85 times faster than the shared-GPU one on equal GPU budgets, while matching their training scores. The paper doesn’t link to a code release.