The results presented in this paper are part of a research project to investigate the possibilities to apply massively parallel architectures for numerical weather forecasting. Within numerical weather forecasting several numerical techniques are used to solve the model equations. This paper compares the performance of implementations on a MasPar system of two techniques, finite difference and spectral, that are adopted in the numerical weather forecasting model HIRLAM. The operational HIRLAM model is based on finite difference methods, while the spectral model is still in a research phase. Also the differences in relative performance of these methods on the MasPar and vector architectures will be discussed.