Behavioral models have been used in the entertainment industry to increase the realism in the simulation of large groups of individuals. Unfortunately, the classical models can be very compute-intensive when very large groups are considered, reducing its applicability in games and other interactive systems. In this article we explore both search space reduction and parallelism to improve the performance of Reynold's Boids model. We propose a methodology that considers self-occlusion (visibility culling) to reduce the number of neighbors and we take advantage the parallelism present in common graphics processor units (GPUs) to allow the simulation of very large groups. We performed different GPU implementations (GPGPU and CUDA); the results show that visibility culling allows significant gains in performance without affecting the model's overall behavior.