We propose belief propagation (BP) based detection algorithms for the Bell labs layered space-time (BLAST) architectures. We first develop a full complexity BP algorithm, and show that the detector achieves a near optimal performance even when the number of receive antennas is smaller than the number of transmit antennas. We also consider three different extensions that provide a good complexity/performance trade-off. Being soft-input-soft-output in nature, we also argue that the proposed detectors are suitable for use in turbo processing which can further enhance the system performance when there is an outer code. In addition to the simulation results, we also study the convergence behavior of the proposed detectors by exploiting the measure of average mutual information.