Laser induced breakdown spectroscopy (LIBS) is proposed to identify tea variety rapidly. LIBS spectra of seven kinds of teas are collected at 200-480 nm wavelength. Two approaches in preprocessing spectra are applied to decrease noise and eliminate disturbance. One is nine-point smoothing (NPS), the other is NPS combined with first derivative (FD). Principal component analysis (PCA) is adopted to reduce the dimensions of processed spectra. Three models like discriminant analysis (DA), radical basic function (RBF) and BP-ANN (multi layer perception, MLP) are selected to discriminate the tea variety. The results demonstrate that the recognition accuracy of tea variety is improved while NPS, FD and PCA are utilized according to priority. And the recognition accuracy of MLP is higher than that of DA and RBF. The recognition accuracy of MLP is 99.6% in training set and 99.1% in test set. It is feasible to select suitable LIBS spectra preprocessing and model construction method to identify tea variety.