A handheld portable device for fruit sugar content was developed based on visible/near-infrared spectral analysis. The device consists of a micro-spectrometer, halogen lamps, OLED screen and microcontroller. The real-time analysis and control software of the microcontroller was written in C language with the help of the Keil 5 development tool. Combined with the spectrum acquisition program written by LabView, the spectra of fruit samples were collected by the developed device. Apples and big peaches were used to explore the detection accuracy of the device and the transfer of the model between two devices (master and slave). The visible-near infrared spectra of the apple and peach were collected in the spectral range of 600~950 nm under laboratory conditions and in the field. The spectral data of calibration set collected by the master device under laboratory conditions were preprocessed by smoothing, maximum normalization, second derivative and other preprocessing methods, followed by the sugar content models developed using partial least squares algorithm for apples and peaches respectively. The models were then imported to the custom software, making it possible for the master device to predict the sugar content of apples or peaches directly. The correlation coefficient and the root mean square error of the prediction set were 0.925, 0.587% and 0.821, 0.613% for apples and peaches, respectively. The models were transferred from the master device to the slave device by using the piecewise direct standardization (PDS) and canonical Correlation Analysis (CCA) algorithm. After comparison, it was found that better model transfer results were achieved based on the CCA algorithm. The correlation coefficient and root mean square error of the prediction set were 0.883, 0.641% and 0.805, 0.626% for apples and peaches, respectively. The model established under laboratory conditions was used to analyze the fruit spectral data collected on the tree, the correlation coefficient and root mean square error of the prediction set were 0.866, 0.741% and 0.816, 0.627% for apples and peaches, respectively. The results showed that the developed device had considerable potential to detect fruit sugar content under lab conditions, and in the field. With the help of the model transfer algorithm, the model can be shared and effectively transferred between different devices. The developed device could meet the demand for rapid, non-destructive, and on-site detection of internal fruit quality.