Background: The use of whole-genome sequence (WGS) data for genomic prediction and association studies is highly desirable because the causal mutations should be present in the data. The sequencing of 935 sheep from a range of breeds provides the opportunity to impute sheep genotyped with single nucleotide polymorphism (SNP) arrays to WGS. This study evaluated the accuracy of imputation from SNP genotypes to WGS using this reference population of 935 sequenced sheep. Results: The accuracy of imputation from the Ovine Infnium® HD BeadChip SNP (~500 k) to WGS was assessed for three target breeds: Merino, Poll Dorset and F1 Border Leicester×Merino. Imputation accuracy was highest for the Poll Dorset breed, although there were more Merino individuals in the sequenced reference population than Poll Dorset individuals. In addition, empirical imputation accuracies were higher (by up to 1.7%) when using larger multi-breed reference populations compared to using a smaller single-breed reference population. The mean accuracy of imputation across target breeds using the Minimac3 or the FImpute software was 0.94. The empirical imputation accuracy varied considerably across the genome; six chromosomes carried regions of one or more Mb with a mean imputation accuracy of <0.7. Imputation accuracy in five variant annotation classes ranged from 0.87 (missense) up to 0.94 (intronic variants), where lower accuracy corresponded to higher proportions of rare alleles. The imputation quality statistic reported from Minimac3 (R²) had a clear positive relationship with the empirical imputation accuracy. Therefore, by first discarding imputed variants with an R² below 0.4, the mean empirical accuracy across target breeds increased to 0.97. Although accuracy of genomic prediction was less affected by filtering on R² in a multi-breed population of sheep with imputed WGS, the genomic heritability clearly tended to be lower when using variants with an R² ≤0.4. Conclusions: The mean imputation accuracy was high for all target breeds and was increased by combining smaller breed sets into a multi-breed reference. We found that the Minimac3 software imputation quality statistic (R²) was a useful indicator of empirical imputation accuracy, enabling removal of very poorly imputed variants before downstream analyses.