Alibaba Qwen launches QwenGyre, a reinforcement learning framework for long-horizon AI agents

Alibaba's Qwen team released QwenGyre this week, a reinforcement learning framework designed to train coding agents on tasks that run for hours and consume close to a million tokens per rollout. The system tackles a practical bottleneck: most existing RL infrastructure assumes short, clean trajectories, but real coding agents produce branching, unpredictable execution histories that break standard training pipelines.

QwenGyre works through unmodified, black-box harnesses. Teams do not need to rewrite agent scaffolding to fit the training loop. Alibaba tested the framework on two models - Qwen 3.6 at 122 billion parameters and the flagship Qwen 3.8 at 2.4 trillion parameters. The 122B runs used 24 to 32 GPU nodes. The 2.4T run scaled to 192 nodes split into four cells of 48. Megatron handles the training; SGLang serves the rollouts.

Two mechanics that handle long-horizon chaos

The system relies on two core mechanisms. The first shifts idle GPUs between rollout generation and training as demand changes, preventing compute from sitting unused during either phase. The second converts the branching, unpredictable histories that long agent runs produce into training samples with fixed, bounded size. This addresses a problem specific to long-horizon work: agent trajectories that fork across hours of tool calls do not fit neatly into standard RL batches.

Claude Code and similar harnesses compact context, spawn sub-agents, and retry on failure. That produces branching trees of execution, not straight lines. Naive implementations duplicate shared work across every branch. Qwen's answer - splitting GPUs into cells with a shared training layout - responds directly to that structural problem.

Performance and timing

On the 2.4T model, QwenGyre completed 48 steps in 75.42 hours. An asynchronous baseline took 134.55 hours, and a colocated approach took 91.47 hours. The evaluation pass rate rose from 52.48% to 58.54%.

Alibaba has not published full comparative benchmarks against standard RL pipelines. The design target - unmodified black-box harnesses at production scale - suggests the framework is built for deployment, not just research demonstrations.

Why this matters for development and operations teams

QwenGyre represents a bet that the real bottleneck in agent training is infrastructure plumbing, not model size. For teams building long-horizon agents, the ability to move GPUs between rollout and training without idling, and to share structure across branching trajectories, could reduce training costs for hour-long, million-token episodes. Developers working on agent orchestration may find the cell-based architecture worth studying, even before independent benchmarks confirm the efficiency claims. Professionals looking to build skills in this area can explore AI Agent Courses and AI for Developers Courses to understand the frameworks shaping production agent systems.