Heterogeneous Feature Fusion Machines (HFFM) is a kernel based logistic regression model which effectively fuses multiple features for visual recognition tasks. However, its batch mode solution suffers inefficiency and poor scalability as common batch algorithm does. In this paper, we developed a novel algorithm based on multiple kernels and group LASSO technique to solve this model, called online HFFM (OLHFFM). The power of the proposed scheme is demonstrated by experiments delivered on public dataset. Moreover, OLHFFM has demonstrated advantages over current state-of-the-art approach as ILK and NORMA in large-scale visual classification.