Qwen releases QwenGyre, an RL framework for ultra-long-horizon agents that reallocates GPUs between rollout and training in real time (community report)

Verified (fact)

According to a Reddit post, Qwen released QwenGyre, a framework for training agents with reinforcement learning on very long, complex tasks (XLong). It has two core design elements. First, an 'Elastic Scheduler' reassigns GPUs between rollout and training without stopping running jobs. Second, a 'Trajectory Processor' converts execution histories that branch through context compression, sub-agents and retries into prefix-sharing trajectory trees, removing duplication and producing size-bounded training samples. Tasks that exceed their time limit are still scored on partial progress. It was reportedly applied to the Qwen3.8-2.4T-A95B model at roughly 700,000 tokens per rollout.

Needs fact-check (claim)

Claimed to be up to 1.85x faster than a colocated setup and up to 1.78x faster than an async setup. Gains on NL2RepoBench were also claimed, but the figures were cut off in the collected excerpt and could not be confirmed.

Why it matters (interpretation)

In agent RL where a single rollout can run to hundreds of thousands of tokens, GPU idle time and handling branching trajectories are the bottlenecks. QwenGyre targets both at once.

Sources

• Community Qwen released QwenGyre - RL framework for training AI agents on extremely long complex tasks Reddit

More about Alibaba

• 2026-10-11 P2 Research Paper Safety & alignmentLanguage models SemiAnalysisAlibabaByteDanceTencentDeepSeekMoonshot AIZhipu AI
  SemiAnalysis: only 3.6% of Chinese AI models have disclosed safety evaluation results, 1.1% before launch (report)

  According to an AI Times report, AI analysis firm SemiAnalysis published an analysis on the 9th (local time) of how Chinese AI...

• 2026-10-10 P2 Model Unconfirmed Language modelsAgents MicrosoftAlibaba
  Microsoft releases 'Microsoft-Decision-1,' a decision model for routing, classification and verification, post-trained from Qwen3.5-9B (community report)

  According to GeekNews, Microsoft has released 'Microsoft-Decision-1,' a model for routing, classification, prioritization, verification,...

• 2026-10-09 P1 Open source Open weights Generative mediaEfficiency & infra AlibabaHugging Face
  Qwen releases open-weight 'Qwen-Image-2.1-Turbo' that generates and edits images in 8 steps (community report)

  According to a Reddit post, the Qwen team released the weights for Qwen-Image-2.1-Turbo on Hugging Face. It is an accelerated checkpoint...