#python

Slime is a high-performance framework for post-training large language models with reinforcement learning (RL). It connects Megatron for fast training and SGLang for data generation, powering top models like GLM-4.7, Qwen3, DeepSeek V3, and Llama 3. You get efficient, flexible RL workflows with customizable data tools, cutting training time and boosting model accuracy for research or production—saving resources while achieving breakthrough results in physics, agents, and code generation.

https://github.com/THUDM/slime