Global optimization of multi-modal functions is a common problem encountered in industrial production. These problems are characterized by numerous local optima, making it difficult to find the global optimum. The Fireworks Algorithm (FWA), as a novel evolutionary algorithm, effectively addresses this issue due to its strong population diversity. However, balancing diversity with convergence efficiency remains a difficult task. Inspired by multi-granularity cognitive computation, this paper proposes a granular-ball based Fireworks Algorithm (GBFWA). By incorporating human cognitive mechanisms into the FWA, the traditional operation design is replaced with a multi-granularity approach. The explosion operations are performed from coarse-to-fine, preserving the algorithm’s diversity while simultaneously enhancing its convergence speed. Additionally, a multi-angle spark mutation operation is introduced, where the difference vectors between the top-ranked sparks of each firework and their current positions collectively guide the evolutionary direction. Extensive experiments on the widely-used CEC2013 and CEC2017 benchmarks demonstrate that these innovations significantly improves the proposed algorithm’s performance, surpassing state-of-the-art FWA variants and outperforming top competitors among other evolutionary algorithms.