Alibaba Qwen unveils QwenGyre: elastic RL for XLong-horizon agents

Alibaba's Qwen team introduced QwenGyre, an online reinforcement learning framework for extreme-long-horizon agent training, targeting rollouts of hours and up to ~1M tokens. On the Qwen 3.8 2.4T model it delivered up to 1.85x speedup over Colocate and lifted NL2RepoBench from 52.5% to 58.5% in 48 steps.

• Up to 1.85x faster than Colocate and 1.78x faster than Async at matched scores
• On Qwen 3.8 2.4T: 48 steps in 75.42 hours vs 91.47 hours for Colocate
• NL2RepoBench rose from 52.48% to 58.54% in 48 steps
• Cap of 5 trajectories per execution cut forward-backward time by 25.2%