SkyRL

Developed by UC Berkeley's Sky Lab and built on VeRL and OpenHands, SkyRL provides a complete RL pipeline for long-horizon, multi-turn tool-use tasks. It treats rollouts as asynchronous services to decouple I/O from GPU training, achieving a 1.55× speedup; this helped boost Qwen3-32B from 24.4% to 39.4% Pass@1 on SWE-Bench Verified with >50% cost reduction. The backend-agnostic design supports SkyRL-train, VeRL, and Tinker, and co-located training/inference enables single-GPU RL. As a fast-moving research project, some experiments yield modest improvements (e.g., +3.6% for a 7B agent), and users need RL and code-agent expertise. It is open-source and free but requires GPU resources. Ideal for researchers prototyping agent RL; less suitable for beginners or production-critical use.