Understanding the three-dimensional (3D) structure and stability of DNA is essential for elucidating its biological functions and advancing structure-based drug design. Here, we present an improved coarse-grained (CG) model for ab initio prediction of DNA folding, integrating a refined electrostatic potential, replica-exchange Monte Carlo simulations, and weighted histogram analysis. The model accurately predicts the 3D structures of DNA with multi-way junctions (e.g., achieving a mean RMSD of ~8.8 Å for top-ranked structures across four DNAs with three- or four-way junctions) from sequence, outperforming existing fragment-assembly and AI-based approaches. The model also reproduces the thermal stability of junctions across diverse sequences and lengths, with predicted melting temperatures deviating by less than 5 °C from experimental values, under both monovalent (Na⁺) and divalent (Mg2⁺) ionic conditions. Furthermore, analysis of the thermal unfolding pathways reveals that the overall stability of multi-way junctions is primarily determined by the relative free energies of key intermediate states. These results provide a robust framework for predicting complex DNA architectures and offer mechanistic insights into DNA folding and function.