Alibaba Qwen Team Introduces QwenGyre Framework for Long-Term AI Agents
Alibaba's Qwen team has announced QwenGyre, a flexible reinforcement learning framework designed for long-running agent training that delivers speedups of up to 85 percent.
The Alibaba Qwen team has introduced QwenGyre, a flexible reinforcement learning framework developed for ultra-long-horizon agent training processes that take hours and process nearly one million tokens.
The Main Purpose of the QwenGyre Framework
Introduced by the Alibaba Qwen team, QwenGyre stands out as an online reinforcement learning system that trains coding agents through unalterable black-box test environments. The framework has been successfully applied to the Qwen 3.8 2.4T model.
Two Core Ideas Behind the System
The QwenGyre architecture is built upon two core ideas: shuttling idle GPUs back and forth between rollout and training processes, and turning branching agent histories into constrained training examples. This way, long-tail operations and branching execution trees are managed.
Flexible Scheduler and Data Transfer
The flexible scheduler divides GPUs into cells, allowing them to share the training layout and make dynamic transitions as executions complete. While memory cache blocks are moved via RDMA, satellite components pull core weights through a transfer engine.
Test Results and Success Rates
Benchmarking tests have proven that the QwenGyre framework offers speedups ranging from 1.21 to 1.85 times higher compared to other systems at matching scores. Performance tests were conducted across various tasks and models.