Qwen introduces QwenGyre, elastic RL for agents whose rollouts approach 1M tokens, with up to 1.85x speedup; the paper puts flagship Qwen 3.8 at 2.4T parameters

Why it matters

【Key points】

On Oct 12 the Qwen team formally introduced QwenGyre, an online RL framework for extreme-long-horizon agents, where one execution can run for hours, span hundreds of interactions and approach 1M tokens. The 12-author paper went on arXiv on Sept 27.

It targets two problems: GPUs sitting idle while long, uneven rollouts finish, and redundant trajectories from branching, retries and compaction. An elastic scheduler moves GPUs between rollout and training without interrupting live executions; a trajectory processor rebuilds branching histories, scores partial progress and deduplicates paths.

【Verifiable】

Paper: https://arxiv.org/abs/2609.33848

Qwen post: https://x.com/Alibaba_Qwen/status/2109633085658808515

On the flagship Qwen 3.8 2.4T with ~700K tokens per rollout, 48 steps raise NL2RepoBench from 52.5% to 58.5%; speedups reach 1.85x over Colocate and 1.78x over Async. The post drew 700+ likes and ~26K views in about 45 minutes. Neither the abstract nor the post links code.

【Our read】

1. The value is a frontier lab showing how it trains its flagship agent. With 700K-token, hours-long rollouts, the bottleneck in agent RL has moved from algorithms to scheduling and dedup.

2. The paper states Qwen 3.8's flagship is 2.4T parameters, a hard number on Qwen's training scale.

3. Evidence is narrow: the 6-point gain is one benchmark, speedups are against two in-house baselines, and without code it is hard to reproduce.