Qwen releases QwenGyre: elastic GPU scheduling speeds long-horizon agent RL training by up to 1.85x
Alibaba_Qwen · x · 2026-10-12
Qwen launched QwenGyre, an elastic RL system for extreme-long-horizon (XLong) agent training, with two core designs:
- Elastic Scheduler: reallocates GPUs between rollout and training as demand shifts without interrupting live executions — up to 1.85x faster than Colocate and 1.78x than Async.
- Trajectory Processor: converts branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples; partial progress after timeouts still gets scored.
Scaled to Qwen3.8-2.4T-A95B with 700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps and cuts end-to-end training time by 44% vs Async. Paper available.
More from coding & agent
- JetBrains study: coding-agent benchmarks should match real users' task flows — JetBrains-Research · 2026-10-12
- Agent Cards and MCP tool descriptions are unverified claims — a team built signed A2A tests — Suncly_com · 2026-10-12
- Voice agent fills CRM fields at 90% confidence — yet payments bounce on wrong account numbers — CoursePopular824 · 2026-10-12
- Dev keeps Codex as code reviewer on top of Claude-generated code — MartinGTobias · 2026-10-12
- How a broken customer ID let a support agent 'resolve' tickets it never read — aiagentguide · 2026-10-12
- OpenEnv Arena offers free GPU compute to train your environment, weights open sourced — ben_burtenshaw · 2026-10-12