Tag
SkyRL v0.4.0 releases with support for GLM 5.3 Flash/5.3, Kimi K2.7, Qwen3.8 and Nemotron 3.5-Lightning, alongside full FP8/MXFP8 RL, Top-P sampler replay, and sharded P2P weight transfer for large-scale post-training.
RollArt presents a disaggregated architecture for large-scale reinforcement learning, demonstrating significant improvements in efficiency and scalability.