Genomic prediction uses a reference population of animals with SNP genotypes and phenotypes to predict the merit of selection candidates that typically do not have observed phenotypes. Accuracy of genomic prediction from models that fitted 50K SNP genotypes was low when selection candidates were from a breed only distantly related with the breeds in the reference population. That accuracy was not improved by increasing SNP density from 50K to a ten-fold higher density using imputation. This indicates that the accuracy of genomic prediction mainly came from family-wise co-segregation information. In this study, a genomic prediction model that fitted genome-wide 100 kilo-bp (Kbp) haplotypes improved accuracy for breeds that were not in the reference population. The haplotype model is a more accurate alternative to the SNP model for genomic prediction when animals of the same breed as the prediction candidates are not available for the reference population.