Leaf concentration in alfalfa is an important factor affecting the nutritive value, forage intake and digestibility. Estimates of leaf concentrations commonly used currently involve a labor intensive process of hand separating leaf and stem fractions. In the present study, a total of 41 artificial alfalfa samples were mixed with different leaf concentrations ranging from 15% to 55%. The object was to develop 3 calibrations for predicting alfalfa leaf concentrations using 15, 25 and 35 calibrated samples by near infrared reflectance spectroscopy. The root mean square error of prediction(RMSEP)was 1.02, 1.97 and 0.51, respectively. External validation had a coefficient of determination (r2) ranging from 0.79 8 to 0.998 9. The ratio of performance to standard deviation (RPD) varied from 2.85 to 25.93. The results showed that 15 samples could develop accurate NIRS model of alfalfa leaf concentrations; the calibration equations got better accuracy with the increase in calibrated samples numbers from 15 to 35.