IT Home, October 12 news: Alibaba's Qwen today officially launched QwenGyre, an online reinforcement learning framework designed for long-horizon agent tasks.
According to official documentation, when agents handle complex tasks, a single execution can last for several hours, involving hundreds of interdependent interactions between the model and the environment, and processing millions of tokens in total. QwenGyre enables such long-duration tasks to become effective training sources through elastic scheduling and trajectory processing, simultaneously enhancing both capability and efficiency.
QwenGyre can divide GPUs into multiple computing units. Once a batch of tasks completes, the freed units are automatically allocated to training, while ongoing tasks remain unaffected—their progress and status are preserved in place, including cached computations that can be fully migrated without restarting. Newly freed units can be directly added to the current training round, and the core unit aggregates all results to complete a parameter update.
IT Home notes that QwenGyre can record the true context for each output generation via Token-in Token-out, organizing it into a trajectory tree based on shared prefixes, and only expanding selected paths as training samples. Additionally, timed-out but partially completed tasks can provide useful feedback through Partial Scoring, avoiding waste of intermediate steps.
Regarding experimental results, the official report states that in the NL2RepoBench online reinforcement learning experiments, each training rollout involves approximately 700,000 tokens. After 48 training steps, the model pass rate increased from 52.5% to 58.5%. Under the same GPU budget, QwenGyre achieves a 1.78x end-to-end speedup compared to Async, with training scores comparable to Async and Colocate. The official also notes that cross-task validation on Qwen3.6-122B demonstrated end-to-end speedups across all three task categories: NL2RepoBench, DeepSWE, and TerminalBench.