Alibaba’s Qwen framework cuts long-horizon agent training time 44%

Key Points

1. Alibaba's Qwen team unveiled QwenGyre, an elastic reinforcement learning framework for agents running hours-long coding tasks.
2. 48 training steps took 75.42 hours versus 134.55 under an asynchronous setup on identical GPU budgets.
3. Cheaper long-horizon reinforcement learning could put multi-hour agent training within reach of smaller compute budgets.

The latest:

Training an AI coding agent on tasks that stretch for hours can now be done in roughly half the time on the same hardware, according to a paper from Alibaba’s Qwen team describing a framework called QwenGyre. The team announced it on 12 October 2026, two weeks after posting the paper to arXiv. Its benchmark runs finished 48 training steps in 75.42 hours.

Details:

• The problem: A single agent task in this regime can run for hours, involve hundreds of consecutive model-environment interactions, and approach one million tokens in a single trajectory, according to the paper. Conventional reinforcement learning pipelines leave GPUs idle while the slowest tasks finish, which is the inefficiency QwenGyre targets.
• Elastic scheduling: The framework splits GPUs into compute units, the paper says. When a batch of tasks finishes, its freed units join the training phase, while unfinished tasks keep running and retain their progress and cached state. The team reported a generation-to-training switch of about 8.52 seconds and the reverse at 3.46 seconds.
• Token-in token-out: The second mechanism logs the actual context of each generated segment and arranges segments in a trajectory tree by shared prefixes, converting only selected paths into training samples. Capping the framework at 5 trajectories per rollout left training scores unchanged while cutting forward and backward pass time by 25.2%.
• Partial scoring: The third mechanism lets tasks that exceed their time limit but are partially completed still produce a usable training signal, rather than being discarded. The paper presents this as a way to recover learning value from the long-running tasks that dominate agentic workloads.
• Headline numbers: On Qwen3.8, a 2.4-trillion-parameter model tested on NL2RepoBench, single trajectories reached roughly 700,000 tokens and success rates rose from 52.48% to 58.54% after 48 steps. Those steps took 75.42 hours with QwenGyre, against 134.55 for Async and 91.47 for Colocate on the same GPU budget.
• Speedups: The reported gains were 1.78 times faster than Async and 1.21 times faster than Colocate on the largest model. On Qwen3.6, a 122-billion-parameter model, speedups ranged from 1.38 to 1.78 times over Async and 1.21 to 1.85 over Colocate across NL2RepoBench, DeepSWE and TerminalBench, peaking on TerminalBench.
• The stack: Experiments ran on multi-node clusters: 24 to 32 nodes for the 122-billion-parameter model and 192 nodes, arranged as 4 cells of 48, for the 2.4-trillion-parameter model. Training used Megatron and generation used SGLang, with a modified GSPO algorithm sampling 16 trajectories per group. The paper does not specify the GPU model.
• Unmodified harnesses: The team designed the framework to train coding agents through operational harnesses as-is, without modification. The harness used in the experiments was Claude Code, version 2.1.220, according to the paper.
• Availability: The paper names no public code repository and no license for the framework, and no release date has been announced. The team’s disclosure so far consists of the arXiv paper, posted 27 September 2026 under number 2609.33848, and the October announcement.

Between the lines:

The efficiency claim rests on a comparison at a fixed GPU budget rather than on fewer chips, so the practical benefit is throughput, not lower hardware requirements. The accuracy gain, from 52.48% to 58.54%, comes from the same 48 steps that define the speedup, meaning the framework is presented as reaching a given result sooner, not a better one.

What’s next

Watch for a code release or license, both absent from the paper, and for independent replication of the 1.78x and 1.85x speedups on hardware the team has not identified.