In near infrared spectral quantitative analysis, many models consider separately each component when modeling sample composition content, disregarding the underlying relatedness among sample compositions. To address this problem, the present paper views modeling each sample composition content as a task, thus one can transform the problem that models simultaneously analyze all sample compositions’ contents to a multi-task learning problem. On the basis of the LS-SVR, a multi-task LS-SVR (MTLS-SVR) model is proposed. Furthermore, an efficient large-scale algorithm is given. The broomcorn samples are taken as experimental material, and corresponding quantitative analysis models are constructed for three sample composition contents (protein, lysine and starch) with LS-SVR, PLS, multiple dependent variables LS-SVR (MLS-SVR) and MTLS-SVR. For the MTLS-SVR model, the average relative errors between actual values and predicted ones for the three sample compositions contents are 1.52%, 3.04% and 1.01%, respectively, and the correlation coefficients are 0.993 1, 0.894 0 and 0.940 6, respectively. Experimental results show MTLS-SVR model outperforms significantly the three others, which verifies the feasibility and efficiency of the MTLS-SVR model.