Rufus-Air is an open, reproducible post-training recipe applied to GLM-4.5-Air-Base (106B-A12B), structured as an eight-stage pipeline spanning SFT, reasoning/coding/instruction-following RL, general-agent, coding-agent, and search-agent stages, and finally RLHF. The paper documents the training data, reward design, and infrastructure needed to reproduce each stage, built entirely from open-source components and public data without new human annotation or a distillation teacher. Key findings: diverse high-quality SFT sets a strong capability floor, difficulty filtering keeps RL prompts productive, and reward reliability should guide stage ordering.
