Temporal graphs provide a powerful framework for modeling dynamic systems in social, biological, and technological domains, where accurate link prediction poses fundamental challenges. Although current approaches successfully model surface-level temporal patterns, they commonly fail to account for underlying structural evolution processes and demonstrate limited reasoning robustness. We present SeqMTG, a Transformer-based framework that redefines temporal link prediction through node sequence generation. SeqMTG introduces an innovative edge serialization technique to capture structural and temporal dependencies of node interactions, coupled with a hybrid spatio-temporal encoding scheme that dynamically models evolving graph connectivity for accurate link prediction. The framework further incorporates multi-token prediction to enhance reasoning capabilities during sequence generation. Experimental evaluations across diverse real-world networks demonstrate consistent superiority over current state-of-the-art methods, validating the effectiveness of our approach in capturing complex temporal graph dynamics while maintaining computational practicality.