slime

Why we included this project

Setting up RL post-training for a large model normally means wiring a trainer, a rollout server, and your own reward logic together by hand. slime keeps that in one codebase: Megatron handles high-throughput training, SGLang generates rollouts, and sampling, reward computation, and environment interaction all flow through a single explicit data path instead of shuttling between disconnected services. When you want to extend the loop with something like tool use or multi-agent workflows, you add it as a data-generation or reward component without touching the training kernel. The project is the RL framework behind the GLM-4.5 through GLM-5.2 releases and also supports the Qwen, DeepSeek, and Llama series, which suggests the full train-to-serve loop has been exercised at real scale. Research groups and platform engineers who want a readable, proven starting point for RL post-training will find that here rather than in an abstract design document.