Alibaba Launches QwenGyre for Faster AI Agent Training

Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning framework built for training extreme-long-horizon coding agents running multi-hour rollouts processing near one million tokens. Tested on models including the Qwen 3.6 122B and the flagship Qwen 3.8 2.4T (A95B per Qwen), QwenGyre optimizes hardware usage by dynamically shifting idle GPUs between rollout and training phases. In benchmarks using the Claude Code 2.1.220 harness with approximately 700K tokens per rollout, the system achieved end-to-end speedups ranging from 1.38x to 1.78x over Async configurations and 1.21x to 1.85x over Colocate setups.

The architecture relies on Megatron for training, SGLang for serving rollouts, and Mooncake's Transfer Engine paired with RDMA for KV-cache migration. Deployed on a 192-node GPU cluster split into four 48-node cells for the 2.4T model, QwenGyre completed 48 training steps on NL2RepoBench in 75.42 hours, outperforming Async at 134.55 hours and Colocate at 91.47 hours. On 122B runs spanning 24 to 32 nodes, it reached a peak 1.85x speedup on TerminalBench, boosted NL2RepoBench pass rates from 52.5% (52.48%) to 58.5% (58.54%) in 48 steps, and registered 1.569x to 1.824x speedups on DeepSWE. Role switching averaged 8.52 seconds from rollout to training and 3.46 seconds back to rollout.

To handle branching sub-agents and context updates efficiently, QwenGyre uses token-in, token-out (TITO) logging to create prefix-shared trajectory trees. The trajectory processor caps samples at 5 per execution—prioritizing main and sub-agent paths—which preserved full accuracy while cutting forward-backward computation time by 25.2%. Training employs an adapted GSPO algorithm with 16 rollouts per group, while partial workspace scoring ensures timed-out executions provide reward signals instead of counting as zero.

For machine learning engineers, QwenGyre proves that unmodified, black-box agent harnesses can undergo online reinforcement learning without severe GPU idle time caused by stragglers. By reallocating compute dynamically and trimming redundant prefix paths, practitioners can significantly reduce training duration and resource costs for complex, repository-scale agent models.