Near-infrared transmittance spectra (740 - 1140 nm) were gathered on single kernels of intact wheat (Triticum aestivum) for the purpose of establishing the feasibility of measuring wheat hardness (i.e., texture) by spectroscopy. Spectra of kernels from a ten-variety hardness standardization set were modeled using multiple linear regression (MLR) on log(1/T) and d2log(1/T)/d(lambda) 2 (up to five terms for each) and partial least squares (PLS) analysis (up to nine factors). Near-infrared diffuse reflectance hardnesses, determined by an official method of the American Association of Cereal Chemists, were the reference values. Single kernel hardness models were then applied to five varieties of wheat excluded in calibration. Results indicated that single kernel hardness by optical measurement of intact kernels is possible, presumably to the extent of the correlation between hardness and vitreousness. However, there is some doubt as to whether intact-kernel transmittance measurements are sensitive enough to measure the biochemical component (presumably, a low-molecular weight protein) that determines hardness. Five-term log(1/T) MLR and eight- factor PLS models provided the best modeling performances. Single kernel hardness models were used to examine kernel-to-kernel variation in hardness. By way of example, when the eight-factor PLS model was applied to the standardization set, Bennett had the least variation (standard deviation of 5.1 NIR-hardness units), and Nugaines had the most (s.d. equals 14.9 NIR-h.u.). Soft wheats tended to have more variation than hard wheats.