SkyRL is a modular, full-stack reinforcement learning library for training large language models, developed by the NovaSky AI team. It provides the training framework and environments needed to run RL on real, long-horizon agent tasks rather than only toy problems. The project has earned more than 2,400 stars on GitHub and was still actively updated in October 2026.
Post-training LLMs with reinforcement learning requires a lot of moving parts: rollout generation, environment interaction, reward computation, and scalable policy optimization all have to work together. SkyRL organizes this into a modular stack so researchers can swap components โ training backend, environments, and agent harnesses โ without rewriting the whole system. It is designed with multi-turn, long-horizon agent training in mind, where a model interacts with tools or environments over many steps before receiving feedback. The documentation site walks through installation, concepts, and examples for getting a training run going. The stack spans both halves of the problem: distributed policy optimization on the training side, and task definitions with reward interfaces on the environment side. Example recipes help new users reproduce reference runs first, then scale up to larger clusters and longer-horizon tasks of their own.
SkyRL is free and open source under the Apache-2.0 license. Training costs are whatever GPU compute you run it on; there is no commercial tier attached to the library.
SkyRL is for RL researchers and ML engineers post-training language models to behave as capable agents. Teams experimenting with reinforcement learning on realistic, multi-step tasks will find its modular stack easier to adapt than monolithic training scripts.