The paper presents Rufus-Air, an open and reproducible post-training recipe for the GLM-4.5-Air-Base model (106B total, 12B active). The recipe is a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. All components use open-source tools and public data, with no new human annotation or distillation teacher.
The authors found that a strong SFT checkpoint builds broad capabilities that later RL stages refine, and that difficulty filtering keeps prompts in a productive learning range. They argue that reward reliability should determine stage order: stages with hard, verifiable rewards run first, and softer judge-based rewards run later. Infrastructure choices—such as token-in/token-out rollouts and a reliable sandbox—are treated as part of the recipe. Rufus-Air outperforms the official GLM-4.5-Air release on every reported benchmark except Arena-Hard v2 Creative Writing, and is competitive with similarly sized open models like INTELLECT-3 and Nemotron-3-Super. The recipe is significant because it provides a complete, reusable account of post-training that other teams can build on.