To determine the tea polyphenols content in Puerh tea by near infrared spectroscopy, genetic algorithm combined with partial least square (GA-PLS) is used to select those wave-numbers carrying information that are highly related to the tea polyphenols of Puerh tea. As a result, 40 wave numbers are selected from the spectral range of 10001～4000 cm-1. Then principal component analysis is applied for these selected wave numbers, and the first two principal components are achieved to input into the extreme learning machine (ELM). Using sigmoidal function as active function, the ELM is trained with the number of hidden-layer neurons varying from 1 to 50. Result shows that it will reach the minimum root mean square error of prediction (RMSEP) set and get the optimal ELM model when there are 13 hidden-layer neurons. The optimal ELM model gives the correlation coefficient of prediction set R2 of 0.9705, with root mean square error of cross-validation (RMSECV) of 1.0109 and RMSEP of 1.6686. This model uses only 40 wave numbers and 2 input-layer neurons, but obtains better performance, compared with the global PLS model with 1557 wave numbers and GA-PLS model with 40 wave numbers. Results show that the tea polyphenols content in Puerh tea can be determined with high precision by using near-infrared spectroscopy combined with genetic algorithm and ELM.