Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents
Alibaba's QwenGyre trains coding agents through hour-long rollouts at 1M tokens—elastic RL for extreme-horizon autonomy.
Why it matters
QwenGyre demonstrates a production RL training system for agents that operate over extended horizons with near-realistic execution windows. This addresses a concrete bottleneck: most agent training assumes short rollouts; QwenGyre targets multi-hour tasks. Practitioners deploying long-running autonomous workflows should track whether this training approach translates to reliability gains in production.
The key facts
7 to knowOnline reinforcement learning framework for agent training
Targets rollouts that run for hours
Processes close to 1M tokens per rollout
Trains coding agents via unmodified, black-box harnesses
Elastic RL framework (resource scaling during training not detailed in excerpt)
Alibaba Qwen team release
Framework name: QwenGyre
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning RL framework for extreme-long-horizon agent training. QwenGyre is an online reinforcement learning system from Alibaba that trains coding agents through unmodified, black-box harnesses. It targets rollouts that run for…