slime Documentation

slime is an LLM post-training framework for RL scaling, providing two core capabilities:

• High-Performance Training: Supports efficient training in various modes by connecting Megatron with SGLang;

• Flexible Data Generation: Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.

slime is the RL-framework behind GLM-4.5 and GLM-4.6. Apart from models from Z.ai, we also supports the following models:

• Qwen3 series (Qwen3Next, Qwen3MoE, Qwen3), Qwen2.5 series;

• DeepSeek V3 series (DeepSeek V3, V3.1, DeepSeek R1);

• Llama 3.

Get Started

• Quick Start
• Usage Guide
• FAQ

Dense

• Qwen3-4B with 8xH100
• GLM4-9B with 8xH100

MoE

• Qwen3-30B-A3B with 8xH100
• GLM-4.5 with 64xH100
• DeepSeek R1 with 128xH100

Advanced Features

• Reproducibility
• Speculative Decoding
• Fault Tolerance
• Supporting Model Architectures Beyond Megatron-LM

Other Usage

• SFT Qwen3-4B-Base
• Search-R1 lite
• Fully Asynchronous Rollout Example
• Retool: from SFT to RL
• Multi-Agent RL

Developer Guide

• Debugging

Hardware Platforms

• AMD

Blogs

• v0.1.0: Redefining High-Performance RL Training Frameworks
• slime: An SGLang-Native Post-Training Framework for RL Scaling