Abstract A parallel master–slave model of the recently proposed cooperative micro-particle swarm optimization approach is introduced. The algorithm is based on the decomposition of the original search space in subspaces of smaller dimension. Each subspace is probed by a subswarm of small size that identifies suboptimal partial solution components. A context vector that serves as repository for the best attained partial solutions of all subswarms is used for the evaluation of the particles. The required modifications to fit the original algorithm within a parallel computation framework are discussed along with their impact on performance. Also, both linear and random allocation of direction components to subswarms are considered to render the algorithm capable of capturing possible correlations among decision variables. The proposed approach is evaluated on two types of computer systems, namely an academic cluster and a desktop multicore system, using a popular test suite. Statistical analysis of the obtained results reveals that, besides the expected run-time superiority of the parallel model, significant improvements in solution quality can also be achieved. Different factors that may affect performance are pointed out, offering intuition on the expected behavior of the parallel model.