Fixed regression model (FRM) analyses that consider only fixed, non-genetic effects to vary over the lactation arecurrently used for genetic evaluation of production traits in South African Holstein. With random regressionmodels (RRM), the random animal and permanent environmental effects are allowed to also vary over thelactation. Hence, RRM can account for an individual component representing changes during the lactation i.e., itspersistency (PERS), enabling selection for more persistent cows. Also, test-day (TD) records used for geneticevaluations come from cows in contrasted production systems. The main ones rely on full pasture (PAST) or atotal mixed ration (TMR), a choice often depending on local average rainfall where herds are situated. TD recordsfrom herds were divided into 2 datasets based on the production system (PAST or TMR). REML was used toanalyze production for each of the first 3 lactations under different multiple-lactation models for milk, butterfatand protein production, as well as butterfat and protein percentage. Various FRM were compared to the currentFRM officially used for genetic evaluation in South Africa (saFRM). A FRM that cumulates different curves overthe lactation for different fixed effects was retained based on results in the PAST dataset and was also applied tothe TMR dataset. This model was then broadened to an alternative RRM (aRRM) combining for each lactation anaverage production and a PERS effect, after which it was compared to the current saFRM under both productionsystems. The aRRM for both PAST and TMR had a better goodness of fit than the current saFRM for all traitsexcept protein percentage. The mean squared error of aRRM was lower for all traits. Generally, aRRM heritabilityestimates (h2) were higher than with the saFRM at the beginning and end of lactation for most traits in PASTwhile being mostly higher during late lactation in TMR. Overall, the h2 in PAST were mostly higher than in TMRfor all traits. Est