What is QwenGyre?
Alibaba's Qwen team introduced QwenGyre, an elastic reinforcement learning (RL) framework built for extreme-long-horizon agent training. It is not a model but a training system: it trains coding agents through unmodified, black-box harnesses. The target is rollouts that run for hours and process close to 1 million tokens.
What problem does it solve?
Two issues dominate at this scale. First, rollouts have long tails: some executions finish very slowly while GPUs sit idle. Second, harnesses like Claude Code compact context and spawn sub-agents, so one execution becomes a branching tree, and expanding every path inflates training cost.
How does the elastic scheduler work?
QwenGyre splits GPUs into cells. Each cell either serves rollouts or hosts a full training replica. A core cell owns the authoritative weights and optimizer state, while satellite cells join training later. As executions finish, cells switch to training. Model calls are rerouted through a proxy so the harness keeps running. Average role switches took 8.52 seconds (rollout to training) and 3.46 seconds (training to rollout).
