It is vital to understand the characteristics of soils and their distribution in space and over time. Spectroscopy in the visible-near-infrared (Vis-NIR) can estimate soil properties (e.g., SOC). Compared with traditional laboratory physical and chemical analysis, spectral technology enables the practical acquisition of soil information rapidly. The development of a soil spectral library (SSL) can provide large amounts of soil data with variability and diversity for empirical calibration. Calibrations derived with these SSLs, however, at the very least, help to improve the robustness of spectroscopic models at regional and local scales due to high soil heterogeneity and model adequateness. Previous studies usually put several target samples into SSL, called spiking; however, the cost-efficiency of spectral techniques was offset more or less. Without spiking samples, we aim to explore the feasibility of developing a local model by constraining the SSL with spectral dissimilarities using classical distance methods. The response between the capacity of the local model with prediction accuracy was also compared and analyzed. In this study, we built a local test set (Test) with the amount of spectral variation from 97 cores, divided by one-tenth of each country from the global soil spectral library (677 cores), and the remaining 580 cores were used as the SSL. We used Euclidean distance (ED), Mahalanobis distance (MD) and Spectral Angle Mapper (SAM) to measure the spectral dissimilarity between Test and SSL and to generate the distance matrix. For each method, nine Local subsets were selected and developed by selecting the spectra of SSL, which were considered similar to the Test. The selection based on the first 0.04%, 0.05%, 0.1%, 0.2%, 0.3%, 0.4%, 0.5%, 1% and 5% of the distance matrix. The statistical models were built to predict SOC concentrations from the spectra by partial least-squares regression. We decomposed the spectra using principal components analysis (PCA) to identify those variables of Local derived from ED, MD and SAM. Our results showed that all the Local models developed by the three distance algorithms without spiking samples still can improve the accuracy compared to the global one, but the inflection points of a sample size of Local with accuracy were significantly different. The SAM considers the waveform and amplitude of the spectrum, so it has more advantages than MD and ED. Its Local, with the first 0.2% ratio, performed the best prediction accuracy, also required the least samples for modeling. We conclude that SAM is more suitable for developing local models from SSL. The first 0.2% of the distance matrix can be used as a reference for the capacity of the local model.