Qwen introduces QwenGyre to improve training for long-running AI agents

Qwen has introduced QwenGyre, a reinforcement learning framework for AI agents that perform lengthy tasks. Its researchers say the system addresses idle computing hardware and duplicated training data during executions that can last hours..

The framework has two main components. Its Elastic Scheduler shifts GPUs between running agents and training models while keeping active executions running. Its Trajectory Processor organizes branching execution histories, evaluates partial progress and removes redundant training paths..

The researchers reported speedups reaching 1.85 times the Colocate baseline and 1.78 times the Async baseline. In a test using Qwen 3.8 2.4T, the NL2RepoBench score increased from 52.5% to 58.5% after 48 training steps. These are results from the team's evaluations, rather than independently established performance guarantees..

The work targets a growing training challenge: agent tasks can involve hundreds of interactions and substantial differences in completion time. That leaves fixed GPU allocations underused while producing overlapping records of an agent's work. The paper describes the framework and experiments but provides no public software release schedule..