In this work the density gradient theory in combination with different Gibbs excess energy (GE) models is used to model the demixing behavior of binary polymer mixtures. The investigated polymer mixtures are Polydimethylsiloxane/Polytetramethyldisiloxanylethylene, Phenylethylether/Polydimethylsiloxane and Boltorn U3000/propanol. As GE-models, the Koningsveld–Kleintjens model and the Lattice Cluster Theory combined with the extended chemical association lattice model were used. So we could show that the density gradient theory combined with different GE-models to consider molecular influences, as association or molecular architecture, can model the demixing of polymer mixtures. The influence parameter required by the density gradient theory was fitted to experimental interfacial tension data. We have shown that the density gradient theory allows modeling the experimentally observed phenomena known from phase equilibria measurements as longer demixing times near the spinodal curve and for shorter tie lines. Moreover, the density gradient theory predicts a change in the interfacial composition near the equilibrium. But this composition change cannot be considered by common sharp interface models assuming a stable interface during the whole separation process.