The accuracy of genotype imputation from a customized low-density to a moderate-density single-nucleotide polymorphism marker panel was tested in Italian Mediterranean buffalo using different sizes of reference populations (RP) and imputation programs. Masked genotypes were imputed using Beagle, minimac and FImpute. The accuracy of predicted genotypes was assessed by the percentage of correctly imputed genotypes. As expected, overall imputation accuracy was better for larger RP sizes. However, increasing the RP size from small to intermediate (i.e., from 42 to 202) resulted in a greater improvement in imputation accuracy compared to increasing the RP size from intermediate to large (i.e., from 202 to 736). The percentage of correctly imputed genotypes was 97.3%, 97.9% and 98.0% for minimac, Beagle and FImpute, respectively, using 736 reference individuals. The accuracy of imputation tended to decrease for high minor-allele frequency of imputed genotypes especially when the RP size was small. The variation in accuracy with increasing RP size was highest for Beagle and lowest for minimac. The accuracy of imputation was larger when the genomic relationship between the reference and the target populations was higher for all three imputation software tools used in this study. The trend of increasing imputation accuracy with genomic relationship was strongest when the RP size was small. Our results indicate that imputation of genotypes from a low to a moderate density panel (i.e., 10K to 90K) is a valid option to reduce the cost of genotyping for genomic evaluation in Italian Mediterranean buffalo.