Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85x
Alibaba’s Qwen team introduced QwenGyre, an elastic reinforcement learning framework designed for extreme-long-horizon agent training. It targets tasks running for hours and processing nearly 1M tokens per rollout, tested on Qwen’s 3.6 122B and 3.8 2.4T models. QwenGyre improves efficiency by dynamically reallocating idle GPUs between rollouts and training, reducing idle time and handling…
Key points
- QwenGyre cuts training time by up to 1.85x over Colocate for long-horizon agent tasks
- Tested on Qwen 3.6 122B (24–32 nodes) and Qwen 3.8 2.4T (192 nodes) with ~700K tokens per rollout
- No public code release, but NL2RepoBench performance rose from 52.5% to 58.5% in 48 steps
The framework achieves 1.38x to 1.78x speedups over Async and 1.21x to 1.85x over Colocate, with a 1.85x end-to-end speedup on TerminalBench. On NL2RepoBench, QwenGyre improved performance from 52.5% to 58.5% in 48 steps. However, no public code release is linked, and performance varies—some runs saw minimal gains due to timeouts. The system relies on multi-node GPU clusters, with the 2.4T model using 192 nodes (4 cells of 48).
The story so far
5 episodes →- Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85xthis story
Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents
MarkTechPost · 12 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic research suggests small AI models may self-replicate on IoT devices · 10 src
- OpenAI says its Navier-Stokes proof does not follow Spanish mathematicians' approach · 1 src
- Parsimony, an open-source benchmark, measures code added by AI models · 1 src
- Meta, Google and NIH invest $1.8B in AI cell simulation project · 1 src
- OpenAI releases 722 math papers from unreleased model · 1 src
Comments
via GitHub Discussions