QwenGyre
An open-source elastic reinforcement learning training framework from the Qwen team, designed for ultra-long-horizon agents. It handles extreme long-horizon agent training through elastic scheduling and prefix-sharing trajectory trees, and supports dynamic GPU reallocation between rollout and training. The framework targets agent workloads whose episodes far exceed conventional training lengths, aiming to sustain both throughput and policy freshness throughout extended training runs.
Timeline (last 90 days)
Alibaba open-sourced a framework called QwenGyre, aimed at solving the problem of GPUs idling while waiting during long-task agent training, using asynchronous scheduling to let agents complete training while executing long tasks.
The core mechanism of QwenGyre is to split the GPU into independent compute units, where units that finish first enter training first, and the state and cache of units that have not finished their tasks are migrated for subsequent continuation, thereby reducing GPU idle time.
QwenGyre targets rollouts of nearly 1M tokens and on the order of several hours.
QwenGyre supports unmodified black-box harnesses, that is, agents run in real, unmodified toolchains, and the training system observes and collects signals from the outside.
The source material did not disclose QwenGyre's specific algorithm details, reward design, or training-benefit data.
QwenGyre still scores partially completed progress after a task times out, so that unfinished branches can also contribute gradient signals.
QwenGyre was validated on the large-scale MoE model Qwen3.8-2.4T-A95B, with a single rollout on the order of about 700,000 tokens.
The Qwen team has open-sourced QwenGyre, an elastic reinforcement learning framework for ultra-long-horizon agent training (XLong agent training).
The link to the QwenGyre paper is arxiv.org/abs/2609.33848.
On the NL2RepoBench benchmark, QwenGyre raised the score from 52.5% to 58.5% in 48 steps, and end-to-end training time was shortened by about 44% compared with the Async scheme.
4 more timeline events
All Facts (14)
The core mechanism of QwenGyre is to split the GPU into independent compute units, where units that finish first enter training first, and the state and cache of units that have not finished their tasks are migrated for subsequent continuation, thereby reducing GPU idle time.
50%UnverifiedQwenGyre targets rollouts of nearly 1M tokens and on the order of several hours.
50%UnverifiedQwenGyre supports unmodified black-box harnesses, that is, agents run in real, unmodified toolchains, and the training system observes and collects signals from the outside.
50%UnverifiedThe source material did not disclose QwenGyre's specific algorithm details, reward design, or training-benefit data.
50%UnverifiedQwenGyre's Trajectory Processor organizes Agent forked execution trajectories into a prefix-shared trajectory tree to reduce sample redundancy and obtain a bounded, deduplicated set of training samples.
50%UnverifiedQwenGyre still scores partially completed progress after a task times out, so that unfinished branches can also contribute gradient signals.
50%UnverifiedQwenGyre was validated on the large-scale MoE model Qwen3.8-2.4T-A95B, with a single rollout on the order of about 700,000 tokens.
50%UnverifiedThe link to the QwenGyre paper is arxiv.org/abs/2609.33848.
50%UnverifiedOn the NL2RepoBench benchmark, QwenGyre raised the score from 52.5% to 58.5% in 48 steps, and end-to-end training time was shortened by about 44% compared with the Async scheme.
50%UnverifiedThe core design of QwenGyre includes two parts: the Elastic Scheduler and the Trajectory Processor.
50%UnverifiedCompared with the Colocate scheduling scheme, QwenGyre brings up to a 1.85× speedup, and compared with the Async scheme it brings a 1.78× speedup.
50%UnverifiedQwenGyre's Elastic Scheduler does not interrupt tasks that are currently executing when reallocating GPUs, avoiding invalidated trajectories.
50%UnverifiedAlibaba open-sourced a framework called QwenGyre, aimed at solving the problem of GPUs idling while waiting during long-task agent training, using asynchronous scheduling to let agents complete training while executing long tasks.
50%UnverifiedThe Qwen team has open-sourced QwenGyre, an elastic reinforcement learning framework for ultra-long-horizon agent training (XLong agent training).
50%