A description is given of the global data assimilation system developed at the National Meteorological Center for treating the heterogeneous global data base. The system is based on updating a nine-level primitive equation prediction model by a local, multivariate, three-dimensional, statistical interpolation method. Updating is performed at 6 h intervals, directly in the prediction model's vertical coordinate. Observations are weighted according to estimates of the observational error of data from different sources. Preliminary performance characteristics are discussed and illustrated by examples.