{"version":1,"type":"story","url":"https://digestai.news/story/alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x","json":"https://digestai.news/story/alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x.json","markdown":"https://digestai.news/story/alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x.md","slug":"alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x","headline":"Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85x","summary":"Alibaba’s Qwen team introduced **QwenGyre**, an elastic reinforcement learning framework designed for extreme-long-horizon agent training. It targets tasks running for hours and processing nearly **1M tokens** per rollout, tested on Qwen’s **3.6 122B** and **3.8 2.4T** models. QwenGyre improves efficiency by dynamically reallocating idle GPUs between rollouts and training, reducing idle time and handling branching agent histories without inflating training costs.\n\nThe framework achieves **1.38x to 1.78x speedups** over Async and **1.21x to 1.85x** over Colocate, with a **1.85x end-to-end speedup** on TerminalBench. On NL2RepoBench, QwenGyre improved performance from **52.5%** to **58.5%** in 48 steps. However, no public code release is linked, and performance varies—some runs saw minimal gains due to timeouts. The system relies on multi-node GPU clusters, with the **2.4T model** using **192 nodes** (4 cells of 48).","keyPoints":["QwenGyre cuts training time by up to 1.85x over Colocate for long-horizon agent tasks","Tested on Qwen 3.6 122B (24–32 nodes) and Qwen 3.8 2.4T (192 nodes) with ~700K tokens per rollout","No public code release, but NL2RepoBench performance rose from 52.5% to 58.5% in 48 steps"],"whyItMatters":"Faster training for long-horizon agents could reduce costs for tasks like codebase generation, but no public tools or benchmarks confirm real-world impact yet.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["Alibaba","Qwen","MarkTechPost"],"models":["Qwen 3.6 122B","Qwen 3.8 2.4T","Claude Code 2.1.220"],"people":["Asif Razzaq"]},"firstPublishedAt":"2026-10-12T21:41:50Z","updatedAt":"2026-10-12T21:41:50Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents","url":"https://marktechpost.com/2026/10/12/alibaba-qwen-introduces-qwengyre-elastic-reinforcement-learning-rl-for-xlong-horizon-agents","publishedAt":"2026-10-12T21:41:50Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":{"title":"AI Audio Arms Race Accelerates with Low‑Cost Real‑Time Voice","url":"https://digestai.news/thread/nari-unveils-qwen3tts-and-qwen3asr-with-low-latency-and-competitive-pricing","storyCount":5},"cite":{"text":"Digest AI, \"Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85x\", 12 October 2026, https://digestai.news/story/alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x","publisher":"Digest AI","title":"Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85x","datePublished":"2026-10-12T21:41:50Z","url":"https://digestai.news/story/alibabas-qwengyre-speeds-up-long-horizon-agent-training-by-up-to-1-85x"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}