Mid-infrared (MIR) reflectance spectroscopy is commonly studied as a rapid and nondestructive method for predictive soil analysis under laboratory conditions. The first objective of this paper is to report an MIR spectral library based on 20,000+ soil samples collected from the United States. The second objective is to assess, using partial least squares regression (PLSR) and artificial neural networks (ANN), the performance of the library to predict 12 physical and chemical soil properties: organic carbon (OC), inorganic carbon (IC), total carbon (TC), total nitrogen (TN), clay, silt, sand, Mehlich-3 extractable phosphorus (P), NH4OAc extractable potassium (K), cation exchange capacity (CEC), total sulfur (TS), and pH. The third objective is to investigate whether the use of auxiliary variables of master horizon (HZ), taxonomic order (TAXON), and land use land cover (LULC) would improve MIR model performance. The results showed that OC, IC, TC, TN and TS were predicted most satisfactorily with R2 > 0.95 and RPD (ratio of performance to deviation) > 5.5. Soil CEC, pH, clay, silt, and sand were also predicted satisfactorily with R2 > 0.75 and RPD > 2.0. P and K were predicted poorly, with R2 < 0.4 and RPD < 1.4. The ANN models generally outperformed PLSR models, except for clay, silt and sand. Using auxiliary variables (HZ, TAXON, and LULC) to develop stratified models generally improved model performance. The HZ-specific models showed the greatest improvements. Using an MIR spectral library for routine soil analysis would positively impact many modern applications where high spatial resolution, quantitative soil data are demanded.