Alibaba Qwen releases QwenGyre, an online RL framework for long-horizon agent tasks
Alibaba's Qwen team released QwenGyre, an online reinforcement learning framework aimed at long-horizon agent tasks, on October 12. According to the company, such tasks can run for hours, involve hundreds of interdependent interactions with the environment and process millions of tokens; QwenGyre splits GPUs into compute units so that units freed when a batch of tasks finishes roll into training while running tasks keep their progress and caches, and it uses Token-in Token-out recording to organize shared-prefix trajectories into a tree, expanding only selected paths as training samples while Partial Scoring salvages feedback from timed-out but partially completed tasks. In the reported NL2RepoBench online RL experiment, with roughly 700,000 tokens per rollout, pass rate rose from 52.5% to 58.5% after 48 training steps, and QwenGyre delivered a 1.78x end-to-end speedup over an Async implementation at the same GPU budget, with cross-task end-to-end speedups on Qwen3.6-122B across NL2RepoBench, DeepSWE and TerminalBench.
Long-horizon agent training is often bottlenecked by the cost of environment interaction and scheduling, so reproducible efficiency gains here would lower the compute barrier to training agents on very long tasks.