Verified project record

THUDM/slime

slime is a lightweight, opinionated reinforcement learning post-training framework that connects Megatron training with SGLang rollout. It serves as the RL framework behind the GLM model family and supports agentic, multi-turn, and custom data generation workflows.

Project overview

It has been validated through GLM model releases, offering an opinionated approach that focuses on the Megatron and SGLang stack for large-scale RL rather than providing a loose collection of disconnected components.

Project type: AI Agent · Model Development · Model Runtime
Use cases: Automation
Deployment: Refer to project documentation
License: Apache-2.0

Best for

AI engineers and researchers who need a lightweight, opinionated RL framework deeply optimized for the Megatron + SGLang training path.
Teams that want a framework proven by frontier model training behind the GLM model family.

Key capabilities

Supports efficient training in various modes by connecting Megatron with SGLang.
Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.
Passes Megatron arguments through directly and exposes installed SGLang arguments with a `--sglang-` prefix.
Supports PD Disaggregation for multi-turn and agentic workloads with different prefill/decode resource needs.
Ships examples for multi-agent rollout, search/RAG-style multi-turn generation, fully-async rollout, and SWE coding-agent RL.

Limitations and risks

The framework optimizes deeply for SGLang as the single rollout backend instead of supporting multiple inference engines.

Getting started

Setup is considered hard because it requires understanding RL, Megatron, and SGLang configuration via command-line arguments, alongside environment setup and data preparation before starting training.

Evidence and sources

GitHub project description: slime is an LLM post-training framework for RL Scaling.
README: slime's design goal is to make these two capabilities reinforce each other without turning the system into a heavy stack of disconnected trainers, rollout services, and agent fram…
README: slime is the RL framework behind GLM-5.2, GLM-5.1, GLM-5, GLM-4.7, GLM-4.6, and GLM-4.5.
README: Lightweight and opinionated: slime focuses deeply on the Megatron + SGLang path used for large-scale RL.
README: Beyond the GLM family, slime also supports: - Qwen series: Qwen3.6, Qwen3.5, Qwen3Next, Qwen3MoE, Qwen3, Qwen2.5; - DeepSeek V3 series: DeepSeek V3, V3.1, DeepSeek R1; - Llama 3.

Reviewed at: 2026-07-18T06:14:49Z