What was released
Alibaba's Qwen team has introduced QwenGyre, a reinforcement learning framework for training AI agents on tasks that run for hours and consume close to a million tokens before any reward arrives. According to Startup Fortune, the system is described in a paper posted to arXiv on September 27 and covered by MarkTechPost on October 12.
The framework trains coding agents through unmodified black-box harnesses, meaning the agent's existing tools do not need to be rebuilt. According to the source, it targets rollouts that run for hours and approach 1 million tokens.
The reported figures
According to the source, QwenGyre cut training time by 44 percent on the 2.4-trillion-parameter Qwen3.8 model: 48 training steps took 75.42 hours instead of 134.55. On NL2RepoBench, a benchmark for repository-level coding tasks, the resulting model's pass rate rose from 52.48 to 58.54 percent. According to the source, Qwen's own comparisons show up to 1.85 times the speedup over a colocated setup and 1.78 times over an asynchronous one; these comparisons come from Qwen itself.
The technical bottleneck
According to the source, the problem with long-running tasks is that GPUs sit largely idle during rollouts while the agent waits on tool calls, file reads and intermediate steps, and the reward only arrives at the end. QwenGyre divides GPU capacity into separate compute units according to the source, so idle units can take on training work without disrupting rollouts still in progress.
Assessment
For German companies wanting to further train their own agents, the notable point is that, according to the source, QwenGyre does not require rebuilding the existing agent harness. Whether the framework will be adopted outside Alibaba's own stack remains open according to the source, which names neither a public code repository nor a license for the paper.
The context is a race among Chinese labs for the agent layer; according to the source, DeepSeek is hiring a team to build coding agents meant to compete with Claude Code. According to the source, whether the measured gains can be reproduced is the decisive question.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
