88Z AI, 13 October — Alibaba has launched QwenGyre, an online reinforcement learning framework designed for ultra-long-horizon agent tasks. The framework reduces GPU idling through elastic scheduling and controls trajectory processing to lower the cost of forked training. Tests on the Qwen3.8-2.4T model show a significant increase in NL2RepoBench pass rates for single rollouts of approximately 700k tokens. Under the same GPU budget, it achieves faster acceleration and greater training time savings compared to Async,
At a glance
- QwenGyre optimizes training efficiency for ultra-long-horizon tasks via elastic scheduling and trajectory processing
- Achieves 1.78x acceleration and saves 44% training time compared to Async under the same GPU budget
The update
Alibaba has officially launched QwenGyre, an online reinforcement learning framework tailored for ultra-long-horizon agent tasks. Designed to address efficiency bottlenecks in long-sequence training, the framework minimizes GPU idling through elastic scheduling and employs trajectory processing to manage the high costs associated with forked training. In practical application, QwenGyre was utilized for training the Qwen3.8-2.4T model. In tasks with a single rollout of approximately 700,000 tokens, the NL2RepoBench pass rate increased from 52.5% to 58.5% after 48 training steps. Compared to the Async method, QwenGyre achieved a 1.78x speedup and saved approximately 44% of training time under the same GPU budget. Furthermore, the framework's effectiveness has been validated across tasks on the 122B model.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 千问大模型Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.