Researchers have published Rufus-Air, an open and reproducible post-training recipe built on GLM-4.5-Air-Base (106B-A12B), according to a paper shared on Hugging Face. The work is framed as a serial pipeline of eight stages rather than a single training run.
The stages, in order, are SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. The authors state that the pipeline progresses from basic to advanced capabilities and from hard, verifiable rewards toward softer, judge-based signals.
The paper documents the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Training relies on open-source components and public data, much of it used as released, and the authors note that no new human annotation or in-house distillation teacher was required.
Four main findings are highlighted: diverse, high-quality SFT establishes a strong capability floor; difficulty filtering keeps RL prompts within a productive learning range; reward reliability offers a practical principle for ordering stages; and infrastructure and engineering choices are part of the recipe rather than an implementation detail.
According to the abstract, Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models. The paper is accompanied on Hugging Face by automated recommendations of related work, including FrogNano, Instella-MoE, ZGCM-1, Miles v0.1, ReToolSQL, T1, and RMSWeb.