An objective of commercial beef cattle crossbreeding programs is to simultaneously optimize use of additive (breed differences) and non-additive (heterosis) effects. A total of 6,794 multi-breed and crossbred beef cattle with phenotype and Illumina BovineSNP50 genotype data were used to predict genomic heterosis for growth and carcass traits by applying two methods assumed to be linearly proportional to heterosis. The methods were: 1) retained heterozygosity predicted from genomic breed fractions (HET1) and 2) deviation of adjusted crossbred phenotype from mid-parent value (HET2). Comparison of methods was based on prediction accuracy from cross-validation. Here, a mutually exclusive random sampling of all crossbred animals (n = 5,327) was performed to form 5-groups replicated 5 times with approximately 1,065 animals per group. In each run within a replicate, one group was assigned as a validation set while the remaining four groups were combined to form the reference set. The phenotype of the animals in the validation set were assumed to be unknown, thus, it resulted in every animal having heterosis values that were predicted without using its own phenotype, allowing their adjusted phenotype to be used for validation. The same approach was used to test the impact of predicted heterosis on accuracy of genomic breeding values (GBV). The results showed positive heterotic effects for growth traits but not for carcass traits which reflect the importance of heterosis for growth traits in beef cattle. Heterosis predicted by HET1 method resulted in less variable estimates that were mostly within the range of estimates generated by HET2. Prediction accuracy was greater for HET2 (0.37 - 0.98) than HET1 (0.34 - 0.43). Proper consideration of heterosis in genomic evaluation models has debatable effects on accuracy of EBV predictions. However, opportunity exists for predicting heterosis, improving accuracy of genomic selection, and consequently optimizing crossbreeding programs in beef cattle.