The thermal conductivity of soil is an important soil property, which affects the soil temperature distribution underground. It has significant practical importance in geotechnical and civil engineering design and construction. Using reasonable means to predict it can effectively solve the problems such as time-consuming and complex process. According to the characteristics of nonlinearity and timing of soil thermal conductivity data, this paper proposed a firefly algorithm (FA) optimization limit learning machine (DELM) prediction model (NMI-FA-DELM) under NMI for soil thermal conductivity prediction. The model first screened the key parameters affecting the soil thermal conductivity by NMI, and took the filtered parameters as the data set. Then the soil thermal conductivity was predicted with the FA-DELM optimized by the firefly algorithm, and the predictive results were compared with those of statistical prediction equations, random forest methods, BP neural network models, DELM models, and SVR (support vector regression) models. The results show that the NMI-FA-DELM model can effectively predict soil thermal conductivity, with corresponding root mean square error, average absolute percentage error, a10 index, and determination coefficient of 0.363, 9.667%, 0.961 and 0.92, respectively. The prediction accuracy of the NMI-FA-DELM model is better than that of other prediction models, and the content of viscous soil and sand has greater influence on the prediction results of soil thermal conductivity. This model can significantly improve the prediction accuracy of soil thermal conductivity and provides important guidance for predicting soil thermal conductivity in practical engineering applications.