Alibaba's Qwen launches QwenGyre to speed up AI agent training

AI agents often run for hours on complex tasks, interacting with their environment hundreds of times and burning millions of tokens. Such long tasks were hard to turn into useful training material before.

Alibaba's Qwen team has launched QwenGyre, an online reinforcement learning framework. It splits GPUs into smaller compute units. Once a task finishes, the free unit jumps into training right away, while running tasks stay untouched and keep their cached data.

The framework logs the real context of every output through Token-in Token-out, builds a trajectory tree by shared prefixes, and only expands chosen paths into training samples. Tasks that time out still offer useful feedback through Partial Scoring.

In NL2RepoBench tests, each rollout used about 700,000 tokens. After 48 training steps, the model's pass rate rose from 52.5% to 58.5%. With the same GPU budget, QwenGyre ran 1.78x faster than Async, and it also sped up three task types on Qwen3.6-122B.