Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning for Extreme-Long-Horizon Agents
By Asif Razzaq — October 12, 2026
Alibaba's Qwen team has introduced QwenGyre, an elastic reinforcement learning (RL) framework designed for extreme-long-horizon agent training. QwenGyre is an online RL system that trains coding agents through unmodified, black-box harnesses. It targets rollouts that run for hours and process close to 1M tokens. The QwenGyre elastic RL framework was applied to Qwen's flagship Qwen 3.8 2.4T model and rests on two core ideas: it dynamically moves idle GPUs between rollout and training, and it converts branching agent histories into bounded training samples.
TL;DR
- Scope: A training framework, not a model. Tested on Qwen 3.6 122B and Qwen 3.8 2.4T (A95B per Qwen), with approximately 700K tokens per rollout.
- Infrastructure: Runs on multi-node GPU clusters. The 122B runs used 24 to 32 nodes; the 2.4T run used 192 nodes (four cells of 48). The GPU model was not disclosed. Megatron handles training, while SGLang serves rollouts.
- Performance: 1.38x to 1.78x faster than Async and 1.21x to 1.85x faster than Colocate at matched scores.
- Best case: 1.85x end-to-end speedup over Colocate on TerminalBench.
- Worst case: Only 1.21x over Colocate on the 2.4T model, where many queries ran near the timeout.
- Bottom line:
- Best result: NL2RepoBench rises from 52.5% to 58.5% in 48 steps.
- Limitation: the paper links to no public code release.
What problem does QwenGyre solve?
The research team defines "XLong-horizon" tasks as agent runs spanning hours and hundreds of model and tool interactions. Examples include generating full repositories from specifications, as in ICML-scale benchmarks, where an agent must plan, write, test, and debug code across many steps. Conventional RL infrastructure breaks down at this scale because rollout generation and policy training compete for the same GPU pool. QwenGyre addresses this by decoupling the two workloads and reallocating compute elastically as demand shifts.
How elastic RL works
QwenGyre's first core idea is elastic GPU scheduling: idle GPUs migrate between rollout generation and policy training based on real-time load. This avoids the static partitioning that leaves hardware underutilized during long tool-calling phases or starves training during heavy generation. SGLang serves rollouts, while Megatron performs training, allowing each backend to scale independently.
The second core idea targets memory and sample efficiency. Branching agent histories—where a single task spawns many divergent trajectories—are compressed into bounded training samples. This keeps the training signal dense without unbounded context growth, which matters when individual rollouts approach 1M tokens.
Benchmark results
On TerminalBench, QwenGyre delivered a 1.85x end-to-end speedup over the Colocate baseline, the strongest result reported. Across tests, it ran 1.38x to 1.78x faster than Async and 1.21x to 1.85x faster than Colocate at matched scores. The narrowest gain—1.21x over Colocate—occurred on the 2.4T model, where many queries ran close to the timeout, limiting the benefit of elastic reallocation.
On NL2RepoBench, the framework improved scores from 52.5% to 58.5% within 48 training steps.
2026 context
QwenGyre arrives as long-horizon agentic coding has become a central benchmark target in 2026. Frameworks that can sustain multi-hour rollouts without GPU waste are increasingly critical as frontier models push past trillion-parameter scale and agents are expected to complete repository-level software tasks autonomously. The absence of a public code release, however, limits reproducibility for the broader research community.
via MarkTechPost