QwenGyre open-sources long-horizon agent RL

QwenGyre targets the real bottleneck in agentic RL: keeping expensive accelerators productive while messy, branching executions remain in flight. The results are compelling, but adoption will depend on reproducibility and access to a runnable implementation.

• –Elastic scheduling lets GPU cells switch between rollout and training as workloads change, avoiding fixed resource partitions and fleet-wide synchronization barriers.
• –Its trajectory processor handles context compaction, sub-agent branches, partial progress, and redundant paths—much closer to production agent behavior than simple append-only loops.
• –On Qwen3.8 2.4T, 48 training steps took 75.42 hours versus 134.55 hours with Async, while NL2RepoBench pass rate rose from 52.48% to 58.54%.
• –The key caveat: current reporting does not identify a public code repository or license, so the release may be more immediately useful as an open systems design than as drop-in infrastructure.