Qwen releases QwenGyre, an elastic RL framework for training AI agents on extremely long complex tasks

Qwen has released QwenGyre, an elastic Reinforcement Learning (RL) framework designed for training AI agents on extreme-long-horizon (XLong) tasks. The system introduces two core components to manage the computational demands of long-context agent training.

Core Components

• Elastic Scheduler: Dynamically reallocates GPUs between rollout and training phases as demand shifts, ensuring live executions are never interrupted. This design delivers up to 1.85x speedup over Colocate and 1.78x over Async methods.
• Trajectory Processor: Converts branching executions—such as compaction, sub-agents, and retries—into prefix-shared trajectory trees. This creates bounded, deduplicated training samples and allows partial progress to be scored even after timeouts.

Performance and Scaling

The framework was scaled to Qwen3.8-2.4T-A95B with approximately 700K tokens per rollout. In evaluations on NL2RepoBench, QwenGyre improved scores from 52.5% to 58.5% in just 48 steps. Additionally, it reduced end-to-end training time by approximately 44% compared to Async baselines.