Alibaba has introduced QwenGyre, an elastic reinforcement learning framework designed to train extreme-long-horizon AI agents. The system successfully addresses the challenges of agent rollouts that run for hours and process close to one million tokens. Tested on flagship models including Qwen 3.8 2.4T, QwenGyre achieved significant speedups ranging from 1.21x to 1.85x faster than traditional setups while raising performance on NL2RepoBench from 52.5% to 58.5%.
The framework operates on multi-node GPU clusters and relies on two core innovations to solve scaling inefficiencies. First, its elastic scheduler dynamically moves idle GPUs between rollout tasks and training phases, utilizing KV-cache block migration over RDMA and Mooncake's Transfer Engine to switch roles in mere seconds. Second, the trajectory processor converts complex, branching agent histories into bounded training samples using token-in-token-out logging, partial scoring for timed-out runs, and strict deduplication to prevent inflated training costs.
This framework matters because training ultra-long-horizon coding agents has traditionally been hindered by severe hardware idling and exploding trajectory tree complexities. By dynamically balancing cluster resources and streamlining gradient updates, QwenGyre paves the way for more efficient and capable autonomous software engineering agents.