QwenGyre
Alibaba's Qwen team introduced QwenGyre, an elastic reinforcement learning framework for training coding agents on hours-long rollouts of close to 1M tokens, up to 1.85x faster than Colocate.
QwenGyre trains coding agents through unmodified black-box harnesses such as Claude Code. An elastic scheduler moves idle GPUs between rollout and training while agents keep running. A trajectory processor turns branching agent histories into prefix-shared trees and keeps at most 5 trajectories per execution. On the Qwen 3.8 2.4T model, 48 steps took 75.42 hours against 134.55 for Async and 91.47 for Colocate.
- Date
- Monday 12 October 2026
- Lab
- Alibaba (Qwen)
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| End-to-end speedup vs Colocate (best case, TerminalBench, Qwen 3.6 122B) | 1.85x 1.21x vs Colocate on the 2.4T model, where many queries ran near the timeout | company |
| Speedup vs Async, NL2RepoBench, Qwen 3.8 2.4T | 1.78x 48 steps in 75.42 h vs 134.55 h | company |
| NL2RepoBench pass rate, Qwen 3.8 2.4T | 52.48% to 58.54% 48 training steps | company |
| Forward-backward time cut from capping 5 trajectories per execution | 25.2% matched uncapped training scores | company |
The paper was reportedly posted to arXiv on September 27. Startup Fortune says open sourced, but MarkTechPost says the paper links no public code and lists the license as not disclosed.
Sources
- marktechpost.com/2026/10/12/alibaba-qwen-introduces-qwengyre-elastic-reinforcement-learnin
- startupfortune.com/alibabas-qwen-team-open-sources-qwengyre-to-train-ai-agents-that-run-fo
This record was checked against its sources on 13 October 2026. How we check