AI Agents
12 October 2026
Alibaba introduces QwenGyre for training long-term AI agents
N
New-ZZZ desk
X @Alibaba_Qwen · 3 hours ago
Alibaba Qwen introduced QwenGyre—a new system designed for training autonomous AI agents. This system utilizes elastic reinforcement learning (RL) for solving tasks with extremely long horizons (XLong).
The main feature of QwenGyre is the implementation of an elastic scheduler. It dynamically reallocates GPU resources between training stages and testing (rollout) depending on the current load. This allows for the efficient training of agents to perform complex, multi-stage tasks that were previously inaccessible to AI.
The system is designed to overcome two key problems associated with training agents in conditions of very long-term planning.
Why it matters
- —The system solves the problem of training agents for very complex, multi-step tasks (XLong).
- —The elastic scheduler improves the efficiency of using expensive GPU computing resources.
- —Improves the capabilities of autonomous AI systems for long-term planning.
Key facts
- The QwenGyre system from Alibaba Qwen is presented.
- Uses elastic reinforcement learning (RL).
- Aimed at training agents with an extremely long horizon (XLong).
- Includes an elastic scheduler for dynamic management of GPU resources.
The full text is in the original source. Here we provide a brief summary and key facts.