slime

@slime_framework

The LLM post-training framework for RL Scaling. github.com/THUDM/slime

Joined September 2025

Following

2,112

Followers

Replies Media

Pinned
slime
@slime_framework
Aug 6
slime v0.3.1 is out! 🚀 Unlike previous releases, this one is not about major new features. As AI agents improve our engineering productivity, we can revisit and remove many compromises across the stack. Highlights: • Lower Megatron offload overhead for colocated training •
Release v0.3.1 · THUDM/slime
From github.com
slime
@slime_framework
5h
slime ❤️ SGLang
SGLang
@sgl_project
9h
Our ecosystem project and the RL framework behind GLM series training, Slime, just open-sourced its deterministic train–rollout alignment path for GLM-5.2. Megatron training and SGLang rollout matched down to a 4096-token logprob MAE of 1.9e-7, with exact zero hidden-state diff
slime
@slime_framework
Aug 11
Ever since RL became a core part of LLM training, train–rollout numerical mismatch has been one of the recurring problems in RL infra. There have been several great open-source efforts toward alignment, but production-scale support still often comes with tradeoffs in model
feat(glm5): align Megatron DeepEP training with SGLang rollout by zhuzilin · Pull Request #2262 ·...
From github.com
slime
@slime_framework
Jul 29
Also the starting point of slime~
Zixuan Li
@ZixuanLi_
Jul 28
Today marks the first anniversary of GLM-4.5 and GLM-4.5 Air. It was a defining return for @Zai_org and our first time connecting with so many open-source enthusiasts and developers. Thank you for being part of this journey. We look to the future with gratitude and confidence.
slime
@slime_framework
Jul 23
slime environment update: - SGLang upgraded to 0.5.15.post1 - TransformerEngine upgraded to 2.16