slime

slime is an LLM training framework for reinforcement learning (RL) scaling, developed by THUDM / Z.ai. It connects Megatron for training with SGLang for rollout generation, and routes training, rollout, reward computation, verifier feedback and environment interaction through a shared training / rollout / Data Buffer path.

It exposes Megatron arguments directly and SGLang arguments via a --sglang- prefix. Custom generation functions, reward functions, verifiers and environments plug in through documented interfaces, which allows multi-turn loops, tool calls and agentic workflows. Advanced options include YAML SGLang server-group configuration, prefill/decode disaggregation, delta weight sync, external rollout engines, fully asynchronous rollout and distillation. The README states it is the RL framework behind GLM-4.5 through GLM-5.3 releases, and a YouTube creator reported a more than 2.3x end-to-end throughput gain for long-horizon RL.

Supported models include the GLM family, Qwen (including Qwen3 MoE and Qwen3-Next variants), DeepSeek (V3, V3.1, R1) and Llama 3. It is a code library for ML engineers and researchers with multi-GPU infrastructure, released under Apache-2.0, with documentation covering debugging, reproducibility, fault tolerance, tracing and profiling.