Slime is an open-source LLM training framework from Tsinghua’s THUDM lab, purpose-built for reinforcement-learning scaling โ the post-training stage that turns a base model into a strong reasoning agent. Released under the Apache-2.0 license, it has collected more than 8,700 stars on GitHub and was actively developed through 2026.
RL post-training couples two very different workloads: high-throughput generation (rollout) and large-scale gradient training. Slime wires them together efficiently, connecting Megatron-LM for training with SGLang for fast rollout inference, so samples flow between the two with minimal waste. It supports mainstream RL algorithms used for reasoning models, handles long-context agentic rollouts, and is the framework behind the training of the lab’s GLM-series reasoning models โ which makes it one of the few RL frameworks proven at frontier-lab scale. The project emphasizes throughput engineering as much as algorithms: asynchronous and partially-asynchronous rollout modes keep expensive GPUs busy while samples are generated, a practical bottleneck that decides whether RL experiments finish in days or in weeks.
Slime is free and open source under the Apache-2.0 license. Training costs are whatever GPU infrastructure you run it on.
Slime is developed openly with configurations close to those used for the lab’s own model releases, including scripts for data preparation and reward modeling, which shortens the path from a research idea to a reproducible training run. Its documentation walks through scaling from a single node to multi-node GPU clusters.
It is for research labs and ML teams doing RL post-training of LLMs โ especially reasoning and agent models โ who need a framework that has already survived real large-scale runs.