The objective of this study was to compare five mathematical functions used for modeling lactation curves and to choose the most suitable one for Polish Holstein-Friesian cows. The data used were 1,944,818; 1,548,700; and 1,081,107 test-day milk yields from 220,487 first, 181,165 second, and 128,774 third lactations. Five models were fitted to the test-day data: the Ali and Schaeffer (ALI), Guo (GUO), and Wilmink (WIL) functions, and thirdorder (LEG3), and fourth-order (LEG4) Legendre polynomials. The milk lactation curves were fitted using a multiple-trait prediction method. Several criteria based on the analysis of residuals were used to compare the models. Of the five models, five-parameter functions (ALI and LEG4) gave the best goodness of fit for standard lactations (305-d), whereas three-parameter functions (GUO and WIL) were the worst. For extended lactations (400-d), the ALI function ensured the most correct prediction.