Alibaba Unveils QwenGyre to Speed Up Agent Training

The system redistributes GPUs in real time and cuts training time by 44% on complex agent tasks, according to researchers.

Training autonomous agents to solve real software engineering tasks requires massive compute clusters, but much of that hardware spends its time waiting. To tackle that bottleneck, researchers at Alibaba Token Hub introduced QwenGyre, an elastic reinforcement learning framework designed specifically for long-horizon agent workflows. The architecture was detailed in a paper published on arXiv and announced on X by @Alibaba_Qwen.

The bottleneck of multi-hour agent runs

Standard language model queries finish within a couple of seconds. Autonomous agents work very differently. When navigating multi-file codebases or running complex test suites, an agent executes hundreds of consecutive actions, evaluates intermediate feedback, and can easily generate up to 1 million tokens during a single run.

In standard online reinforcement learning setups, that asymmetry causes severe cluster inefficiencies. Hardware is typically divided into fixed pools for inference (rollouts) and model training. When complicated tasks take hours to resolve, GPUs assigned to gradient updates sit idle for long stretches, blocked until the slowest stragglers cross the finish line.

How QwenGyre shifts GPUs without stalling tasks

The framework tackles this imbalance with an elastic scheduler. Instead of locking GPUs into rigid roles, the architecture segments cluster capacity into independent cells and dynamically reallocates idle nodes to model training on the fly. When a multi-hour rollout is still running, the system reroutes incoming requests and migrates active key-value caches across physical servers over high-speed RDMA networks without aborting the ongoing task.

Alongside elastic scheduling, the team built a dedicated trajectory processor. Autonomous agents stumble, retry, and call external tools, leaving sprawling execution logs with heavy data duplication. QwenGyre structures these runs into shared-prefix trees, pruning redundant context from cluster memory. The tree structure also enables the system to score partial progress when an agent run hits an execution timeout. For developers, the framework interacts with existing test harnesses and scaffolds as unmodified black boxes, requiring no rewrites to the agent’s tool layer.

Reported benchmarks and practical caveats

In internal tests run on the 2.4-trillion-parameter Qwen 3.8 model (Qwen3.8-2.4T-A95B) with rollouts averaging around 700,000 tokens, Alibaba reported that QwenGyre completed 48 reinforcement learning steps in 75.42 hours. A standard asynchronous baseline required 134.55 hours for the same workload. That 44% drop in training time yielded a 1.78x speedup over asynchronous pipelines and 1.85x over colocate setups. On the software engineering benchmark NL2RepoBench, task pass rates climbed from 52.48% to 58.54%.

Those numbers require healthy skepticism. Every metric was generated inside Alibaba’s private data centers, and external teams have not verified network overhead or real-world stability. In addition, several tech outlets claimed Alibaba open-sourced the platform, but that remains an unverified RUMOR: the research paper presents the system design and empirical formulas, but the team has not released a public GitHub repository or an open-source license.

Sources

• arXiv:2609.33848 - Technical paper “QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents” (Alibaba Token Hub)
• X (@Alibaba_Qwen) - Official announcement by the Qwen team
• Gate.com - Analysis of GPU scheduling and elastic reinforcement learning
• Startup Fortune - Industry report on multi-hour agent training infrastructure