Advanced differentiated thyroid cancer (DTC) is characterized by limited therapeutic options and unfavorable prognosis. To address this, we conduct proteogenomic analysis of 113 advanced DTCs, identifying three molecularly distinct subtypes: canonical, stromal, and immunogenic. These subtypes exhibit differences in driver mutations, histopathological features, and clinical outcomes. Based on their unique biology, we suggest distinct therapeutic strategies for each subtype. To facilitate clinical application, we develop a machine learning classifier that accurately predicts these subtypes using routinely available gene mutation and digital pathology data. The biological relevance of this classification is further confirmed in an independent cohort analyzed by single-cell and spatial transcriptomics. Moreover, analysis of a real-world cohort of patients receiving various systemic therapies provides preliminary clinical evidence supporting the potential utility of this subtyping framework for informing treatment decisions. Collectively, this study provides a rationale and a practical tool for future exploration of personalized treatment in advanced DTC.