Alibaba's Qwen team officially released QwenGyre on October 12, an online reinforcement learning (RL) framework designed for ultra-long-horizon agent tasks, marking a new step in the training of agents that work for hours at a stretch.
When agents take on complex work, a single execution often lasts hours, involving hundreds of interdependent interactions between model and environment and cumulative processing of millions of tokens. QwenGyre makes such extreme long-horizon tasks an effective training source through elastic scheduling and trajectory processing.
The framework divides GPUs into compute cells: as each batch of tasks finishes, freed cells switch to training while still-running tasks remain undisturbed, with progress, state, and even previously computed caches migrated wholesale rather than recomputed. On trajectory processing, QwenGyre records the true context of each output segment via token-in token-out, organizes shared prefixes into trajectory trees, and expands only selected paths into training samples. Timed-out but partially completed tasks can still provide useful feedback through partial scoring.
In benchmarks, on the NL2RepoBench online RL experiment using tasks of roughly 700,000 tokens per rollout, 48 training steps lifted the model pass rate from 52.5% to 58.5%. Under the same GPU budget, QwenGyre achieved a 1.78x end-to-end speedup over asynchronous implementations, with training scores close to Async and Colocate approaches. Cross-task validation on Qwen3.6-122B showed end-to-end speedups across NL2RepoBench, DeepSWE, and TerminalBench.
