In this article, a large-load coreless magnetic levitation gravity compensator is analyzed and improved in aspects of model improvement and parameter optimization. Composed of three layers of 2-D permanent magnet arrays, the magnetic levitation gravity compensator can generate a large levitation force and be used as the suspension support in the ground test of large space optical equipment. To increase the model accuracy, the 3D analytical model which only considers the operating point of the initial position of PM is improved. By considering the actual operating points of the PM at all positions, the model error can be reduced, which is verified by results of 3D finite element model. The optimal parameters of the magnetic suspension gravity compensator are obtained using genetic algorithm and a better performance is verified by the improved model as well as the 3D finite element method.