Co-clustering is a data mining technique that aims at identifying the underlying structure between the rows and the columns of a data matrix in the form of homogeneous blocks. It has many real world applications, however many current co-clustering algorithms are not suited on large data sets. One of the successfully used approach to co-cluster large data sets is the MODL coclustering method that optimizes a criterion based on a regularized likelihood. However, difficulties are encountered with huge data sets. In this paper, we present a new two-level co-clustering algorithm, given the MODL criterion allowing to efficiently deal with very large data sets that does not fit in memory. Our experiments, on both simulated and real world data, show that the proposed approach dramatically reduces the computation time without significantly decreasing the quality of the co-clustering solution.