The Slime RL framework has been officially open-sourced by THUDM, offering developers the exact post-training stack used for Zhipu AI’s GLM-5.2 model. This release allows researchers to replicate the parallel training loops that combined multiple expert models into a single system. The tool is designed to handle complex, long-horizon agent tasks efficiently.
What is the Slime RL framework?
The tool serves as a specialized post-training system for reinforcement learning scaling. Instead of relying on disconnected trainers and rollout services, it integrates training, data generation, and reward computation. This integration occurs within a single, unified path. Developers can access the code directly through the official GitHub repository.
Architecture and Core Components
The system consists of three core modules. First, the training module handles the main learning process using Megatron. Second, the rollout module generates new data using SGLang. Finally, the data buffer manages prompt initialization and custom strategies. The Slime RL framework supports both a colocated, synchronous mode and a disaggregated, asynchronous mode. This flexibility is detailed in the LMSYS Org blog post.
Training GLM-5.2 in Two Days
The efficiency of this system is demonstrated by its production pedigree. The entire post-training phase of GLM-5.2 took approximately two days on this platform. This speed is remarkable given that GLM-5.2 is a sparse Mixture-of-Experts model with 750 billion parameters. It features 40 billion active parameters per token and a 1-million-token context window.
Market Impact and Open-Source Access
The release has already influenced the global artificial intelligence sector. According to a South China Morning Post report, the stock of the associated firm surged significantly following the open-source announcement. By releasing the Slime RL framework, THUDM provides a highly reliable infrastructure for global AI development. Users can deploy the model locally using various software tools like SGLang and vLLM.
