A new RBF neural networks model for mineral resource potential mapping is proposed in this paper. For the purpose of applying this new model, a three-step procedure is needed as follows: first step is to get training samples from the study area; second step is to abstract the structure of spatial information of training samples and then to construct a RBF networks; and last step is to generate the distributive map of mineral resource potentials. In the paper, the model was employed to predict multi-metallogenetic prospecting targets in the area from Duolanasayi to Ashele in the northern Xinjiang. The predicted targets by the model were compared with the C-F model. The two model results are very similar to each other, suggesting that the new model is effective and practical.