AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuning code

SkyRL

NovaSky
open source / Overall score: 2.5

SkyRL is a modular reinforcement-learning library for post-training language models and agents. Its unified skyrl package runs PPO, GRPO and related algorithms over FSDP or Megatron training with vLLM or SGLang rollout, and implements the Tinker API so Tinker training scripts run on local GPUs. Alongside it ship skyrl-gym, a set of tool-use environments, and skyrl-agent for long-horizon agents. NovaSky at the Berkeley Sky Computing Lab builds it with Anyscale.

SkyRL tx, the Tinker API backend, and the earlier skyrl-train trainer were separate packages until both moved into the unified skyrl package; their old PyPI names are declared beside it.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI, root LICENSE)+MIT(OSI, declared by the skyrl-gym package)
source
public(NovaSky-AI/SkyRL)
core features withheld
no — the tree at 2c54cef carries one root Apache-2.0 LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path. The project is developed at the Berkeley Sky Computing Lab, and Anyscale runs the same open library on its paid managed platform rather than selling an edition of it

SkyRL is Apache-2.0 licensed, and its public repository holds the trainer, the Tinker backend, the environments and the agent layer, so it builds and runs from published source with no account. Anyscale offers SkyRL on its paid managed platform, but that is a separate service running the same open library, and nothing in SkyRL is held back for it.

Adoption

1 medium confidence
1.0

Adoption is the trailing month of PyPI downloads of skyrl, the package that now holds the trainer and the Tinker backend; the retired skyrl-train and skyrl-tx packages and the skyrl-gym environment library are left out. Installs from a git checkout, which the installation guide uses, and from the project's Docker images are not counted, so the figure is a lower bound on use.

Capability

4 medium confidence
4.0

SkyRL covers a broad RL post-training stack: PPO, GRPO and several newer policy-optimization variants over FSDP or Megatron, vLLM or SGLang rollout, fully asynchronous training, and a local implementation of the Tinker API. It sits a step below verl, which documents a wider algorithm catalog and has been used to train frontier reasoning models.

Verified 2026-10-19, except some scores, still awaiting re-confirmation