<p>Large-scale landslides triggered by intense rainfall often experience prolonged deformation and complex mechanical processes, posing significant challenges for early warning. Developing a reliable early warning model (EWM) is crucial for mitigating losses caused by landslides. This study uses the large-scale Tanjiawan landslide as a case study, integrating 15 years of manual observations, two years of real-time monitoring, and extensive field investigations to analyze its spatiotemporal evolution. The “one rainfall process” concept was introduced to classify rainfall events influencing landslide deformation into antecedent and current modes. In this study, rainfall–displacement kinematic features are defined as the quantifiable variations in displacement rate, incremental rate of displacement, and improved tangential angle of the landslide in response to rainfall forcing. Displacement rate and effective rainfall were used to establish threshold relationships and a progressive correlation ratio criterion. The improved tangential angle was further incorporated to identify deformation stages, leading to the development of a four-level dynamic EWM. The landslide exhibits a periodic movement pattern strongly correlated with continuous rainfall, and its displacement curves display distinct step-like characteristics. Based on the spatial distribution of surface cracks and variations in displacement, separate early warning models were established for the three primary deformation zones (Ⅰ, Ⅱ, and Ⅲ). The model demonstrated strong applicability for early warning of landslides with step-like deformation patterns and showed good agreement with field observations. The zoned EWM substantially improves the accuracy of landslide early warning. Analysis of the deformation response characteristics of the Tanjiawan landslide provides deeper insights into the dynamics of large-scale landslides and offers valuable guidance for landslide monitoring and early warning practices.</p>