Alibaba's Qwen team announced on Monday, October 12, 2026, QwenGyre, an elastic online reinforcement learning (RL) framework specifically designed to train artificial intelligence agents on ultra-long-horizon (XLong-Horizon) tasks.
What is QwenGyre?
Modern AI agents, especially in software engineering, often execute tasks lasting hours, involving hundreds of interactions between the model and the environment, processing hundreds of thousands to millions of tokens. QwenGyre addresses two main problems: GPU idleness due to variable execution times and redundancy in complex trajectories with branching, sub-agents, and retries.
The framework has two core components. The Elastic Scheduler dynamically reallocates GPUs between the rollout phase (task execution) and training without interrupting ongoing executions. When a task finishes, the freed unit immediately joins training, preserving progress, cache, and state. The Trajectory Processor organizes execution histories into prefix-shared trajectory trees, deduplicating samples and enabling partial scoring even for timed-out tasks.
Performance results
Tested on models such as Qwen 3.8 2.4T (with around 700,000 tokens per rollout), QwenGyre completed 48 training steps in 75.42 hours, compared to 134.55 hours for the traditional Async method — a reduction of about 44%. Relative to Colocate approaches, it achieved up to 1.85x end-to-end acceleration.
On the NL2RepoBench, the pass rate rose from 52.5% to 58.5% after 48 steps. Similar acceleration results were observed on Qwen 3.6 122B across tasks like DeepSWE and TerminalBench. The official paper is available on arXiv under number 2609.33848. More details were shared by the team on X and reported by outlets such as IT Home.
Why it matters
As AI agents take on increasingly complex and long-duration tasks, efficient training frameworks are crucial for scaling coding and real-world problem-solving capabilities. QwenGyre demonstrates that it is possible to train these agents practically and efficiently on GPU clusters, paving the way for more capable agents in code repositories and real software environments. This represents an important advance in long-horizon AI agent research, especially for large-scale software engineering tasks.
Primary source: IT Home and official announcement by the Qwen team on X. Paper: arXiv:2609.33848.
By GeekikiBot