From Startup Fortune · · 1m 20s
Alibaba Opens QwenGyre for Long-Running AI Agents
Alibaba's open-source QwenGyre framework trains AI agents to run for hours without breaking.
In brief
Alibaba's open-source QwenGyre framework trains AI agents to run for hours without breaking. Alibaba released QwenGyre, an open-source reinforcement learning framework that trains AI agents to run for hours, cutting training time by 44% and improving coding task pass rates. Originally reported by Startup Fortune.
Alibaba Releases QwenGyre Framework
Alibaba's Qwen team published QwenGyre, a reinforcement learning framework for training AI agents that run for hours.
QwenGyre Targets Long Tasks
The framework trains coding agents through unmodified black-box harnesses, targeting rollouts that run for hours and approach 1 million tokens.
Training Time Reduced by 44%
QwenGyre cut training time by 44%, completing 48 training steps in 75.42 hours instead of 134.55.
Improved Coding Task Pass Rate
On NL2RepoBench, the model's pass rate climbed from 52.48% to 58.54%.
Elastic Scheduling Fix
QwenGyre divides GPU capacity into compute units to share hardware dynamically during training.
Alibaba's 10 Trillion Parameter Plan
Alibaba CEO Eddie Wu plans to train an AI model with 5 to 10 trillion parameters by 2032.
Open Infrastructure for Teams
QwenGyre trains through existing agent harnesses without requiring re-architecting.
QwenGyre Adoption Uncertain
Whether QwenGyre gets adopted outside Alibaba remains an open question.
What matters most
Alibaba released QwenGyre, an open-source reinforcement learning framework that trains AI agents to run for hours, cutting training time by 44% and improving coding task pass rates.
