The traditional multivariate analysis method is the main method for quantitative modeling of LIBS spectral datasets, but the input dimension of the spectrum is relatively high. Reducing the dimension of the spectrum and extracting the characteristic spectral line in advance is needed by many algorithms, which results in partial loss of information and affects the accuracy. Aiming at this issue, a quantitative modeling method based on deep convolutional neural network inception is introduced, and the conventional 2D convolutional network is transformed into 1D form to realize full spectrum input and feature extraction of spectral information. Not only there is no need to reduce the dimension of the original spectrum in this method, but also it omits other preprocessing operations such as filtering. Through many experiments, when the number of training is 2 000, it has a good prediction result with no obvious overfitting phenomenon. Its average coefficient of determination (R²) is 0.957 9, and its root mean square error is reduced to 61.69% of those by Partial Least Squares Regression (PLSR). Compared with PLSR and the AlexNet deep learning method the proposed method both gets better results.