Rufus-Air Open Post-Training Recipe for 106B GLM-4.5-Air
September 23, 2026
Rufus-Air provides an eight-stage post-training pipeline for the GLM-4.5-Air-Base (106B-A12B) model, covering SFT, Reasoning RL, and specialized agent training. The recipe uses public data and open-source components to build advanced capabilities without new human annotations.
HOW THIS AFFECTS YOU
●
builderYou can use this documented pipeline to improve the reasoning and agentic capabilities of large-scale base models.
●
researcherYou can replicate or build upon this multi-stage RL and SFT framework for large-scale model optimization.