Alibaba's Qwen team unveils QwenGyre, a training system for AI agents on tasks that run for hours, claiming up to 1.85 times faster reinforcement learning
Introducing QwenGyre, bringing elastic RL to extreme-long-horizon (XLong) agent training. ✨ It launches with two core designs: 🥳 - Elastic Scheduler: reallocates GPUs between rollout & training as demand shifts—live executions never interrupted. Up to 1.85× speedup over Colocate, 1.78× over Async. - Trajectory Processor: turns branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples. Partial progress after timeout still gets scored. Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps, cutting end-to-end training time by ~44% vs Async. 🔗 Read the paper:
Context
Training agents on very long tasks such as building whole software repositories wastes a lot of GPU time, so systems that cut that waste make longer horizon agents cheaper to build.
QwenGyre is described in an arXiv paper from Alibaba Group with researchers from USTC and Tsinghua University. It targets online reinforcement learning, where a model improves by attempting tasks and being scored. All results are the authors' own and have not been independently reproduced.
Deep dive
On October 12, Alibaba_Qwen posted that QwenGyre brings elastic reinforcement learning to extreme long horizon agent training.
According to the paper, a single agent run in this regime can span hours, hundreds of interactions and nearly 1 million tokens. The authors say two problems follow. Run times vary widely, so GPUs sit idle waiting for the slowest runs, and agent tools that compact context, call sub agents and retry create branching histories with large redundancy.
QwenGyre's elastic scheduler moves groups of GPUs between generating rollouts and training as demand shifts, without interrupting live runs. Its trajectory processor turns branching runs into trees with shared prefixes, scores partial progress after a timeout and caps the number of training samples per run.
The paper reports speedups of up to 1.85 times over a colocated baseline and 1.78 times over an asynchronous one. On the flagship Qwen 3.8 2.4T model with about 700,000 tokens per rollout, it says NL2RepoBench rose from 52.5% to 58.5% in 48 steps, and 48 steps took 75.42 hours versus 134.55 hours for the asynchronous setup. The experiments ran agents through Claude Code as the harness.
What to expect next
Watch whether Alibaba releases QwenGyre code, since the post links only the paper, and whether other labs adopt elastic GPU allocation for long agent training.
Also watch whether the gains on Qwen 3.8 2.4T show up in future Qwen releases, particularly on coding benchmarks such as NL2RepoBench, and how other research groups' results compare at smaller GPU budgets, which the paper says it did not test.