In the last few years, geometric semantic genetic programming has incremented its popularity, obtaining interesting results on several real life applications. Nevertheless, the large size of the solutions generated by geometric semantic genetic programming is still an issue, in particular for those applications in which reading and interpreting the final solution is desirable. In this paper, we introduce a new parallel and distributed genetic programming system, with the objective of mitigating this drawback. The proposed system (called MPHGP, which stands for Multi-Population Hybrid Genetic Programming) is composed by two subpopulations, one of which runs geometric semantic genetic programming, while the other runs a standard multi-objective genetic programming algorithm that optimizes, at the same time, training error and the size of the solutions. The two subpopulations evolve independently and in parallel, exchanging individuals at prefixed synchronization instants. The presented experimental results, obtained on five real-life symbolic regression applications, suggest that MPHGP is able to find solutions that are comparable, or even better, than the ones found by geometric semantic genetic programming, both on training and on unseen testing data. At the same time, MPHGP is also able to find solutions that are significantly smaller than the ones found by geometric semantic genetic programming.