Genomic selection uses molecular and pedigree information to accurately estimate genomic breeding values of animals from birth for traits in selection. Recent research in phenomic selection in plant production is opening up new opportunities in animal breeding. The approach of phenomic selection has been little studied in animal production. Here, we evaluate the efficiency of phenomic selection to estimate the phenomic values of phenotype-free females using mid-infrared spectral (MIRS) data from their milk samples. The phenotypes of 1,531 first-lactation French Lacaune dairy ewes were considered for traits included classically in the breeding goals, such as milk production and functional traits (SCS and udder type traits). The inclusion of standardized raw MIRS data instead of SNPs led to very low phenomic predictive abilities for udder type traits (Pearson correlations between phenotype and phenomic values from −0.08 to 0.07). For milk production traits, the phenomic predictions were superior to the genomic ones, in particular for lactation SCS (LSCS), with a predictive ability at 0.49 instead of 0.04. Overall, random regression-BLUP and Bayesian reproducing kernel Hilbert space methods gave equivalent results on phenomic predictions across all traits, with no impact from spectral data preprocessing. Finally, the efficiency of the combination of SNPs and milk MIRS in prediction models was low (average +3.8% for milk production and LSCS traits). Phenomic predictions could open up new prospects especially for the selection of nongenotyped females.