QwenGyre: elastic RL framework for training xLong-horizon agents
Radar: the QwenGyre framework, described by the Qwen team at Alibaba, reallocates GPUs between rollout and training and groups branched agent runs into trajectory trees, with a reported speedup of up to 1.85x over Colocate and about 44% less training time.
Scope and data: according to the briefing dated October 12, 2026, the Qwen team at Alibaba presents QwenGyre, an elastic reinforcement learning framework for xLong-horizon agents. It reallocates GPUs between the rollout stage and the training stage and groups branched agent executions into trajectory trees with shared prefixes. The author reports a speedup of up to 1.85x over Colocate and about 44% less training time. The text says the goal is to keep GPUs busy and deduplicate samples when RL rollouts for long-horizon agents last hours. The summary is truncated, so methods, hardware, benchmarks and experimental conditions are not included here.
Relevance and verification: this is training infrastructure research. Rota Nacional does not offer RL training and does not solve the GPU allocation problem described. If you use AI to study the material, do not paste logs, agent traces or internal datasets containing names, CPF numbers, e-mails or phone numbers into prompts; use anonymized samples. To check the claim, consult the original reference listed in the source and the authors' published material, and verify the baseline used for the speedup.