QwenGyre: 48 steps of 2.4T agent RL in 75.42 hours, against 134.55 on an async baseline
QwenGyre is an online RL framework for agent executions that run for hours and consume about a million tokens per rollout. It reallocates GPUs between rollout and training while live executions keep running, reconstructs branching execution histories as prefix-shared trajectory trees, and scores partially completed work instead of discarding it. The Qwen team reports 48 training steps on Qwen 3.8 2.4T in 75.42 hours against 134.55 for Async and 91.47 for Colocate, with the evaluation pass rate moving from 52.48% to 58.54%.