Blind separation of underdetermined instantaneous mixtures is a popular solution to inverse problems encountered in audio or biomedical applications where the number of sources exceeds the number of sensors. There are two non-equivalent tasks: to identify the mixing matrix and to separate the original sources. In this paper, we focus on the latter task by proposing a novel beamformer that minimizes the theoretical mean square error distance between the separated and original signals. The beamformer has two stages: one for the estimation of signals and one for their refinement. Within the former stage, the signals are assumed to be random and locally stationary, while the latter stage is based on a semi-deterministic model. The experiments prove superior performance of the proposed method compared to conventional MMSE beamforming.