<p>Exotic maize germplasm and derivation have formulated new heterotic groups in China. The breeding values are urgent to evaluate for better application. Unlike GWAS (genome-wide association study), genomic prediction (GP) can predict breeding values using all the genomic markers jointly rather than testing the significance of each of them. A panel of 636 exotic maize lines derived from the national project was genotyped and crossed to two testers Jing2416 and Z58. The test crosses were evaluated in 2017 and 2018 in two sites. The mean performance of two test crosses for each Line was used to train a whole GP model. Five-fold cross-validation was performed to assess the prediction accuracies of the GP models for all traits in the same population. Meanwhile, the tester GP model of each type test crosses for one tester was also constructed. The result indicated that the accuracy of prediction for all the traits ranged from 0.36 to 0.56 in the whole GP model. The accuracy of the ear width was highest at 0.56; plant height was second at 0.53. The forecast of grain yield was 0.49 lower than the ear width and plant height. The prediction accuracy of Jing2416 model was always above that of the Z58 model, the whole model in the middle for most traits. The major reason was that the genetic relationship of Jing2416 with the training population was far more. Thus, more consanguinity ties of tester should be chosen with the training-validation population. The prediction accuracy of the whole model was always more than that of the Jinan model and Xinxiang model for all the traits. It underlines that the prediction model based on multi-environments had a better forecast result. The single-environment phenotypic value had a lower prediction effect.</p>