A least square smoothing (LSS) approach is presented for the blind estimation of multi-input multiple-output finite impulse response system. By exploiting the isomorphic relation between the input and output subspaces and the code sequences, this geometrical approach identifies the channel from a specially formed least squares smoothing error of the channel output. LSS has the finite sample convergence property, i.e., in the absence of noise, the channel is perfectly estimated with only a finite number of data samples. Referred to as the adaptive least squares smoothing algorithm, the adaptive implementation has a fast convergence rate. A-LSS is order recursive, and can be implemented using lattice filter and systolic array. It has the advantage that, when the channel order varies, channel estimates can be obtained without structural change of the implementation. For uncorrelated input sequence, the proposed algorithm performs direct deconvolution as a byproduct.