Granular-ball computing is an important model developed in the field of granular computing in recent years. This method uses granular-balls to represent and cover the sample space, and learns based on granular balls. In the current unsupervised granular-ball partitioning method, the DM (Distribution measure) value or a fixed value is mainly used to measure the quality of the granular-ball. The former can clearly represent the data distribution, but resulting in a larger number of granular-balls obtained. The latter may not clearly express the data distribution. We hope to obtain fewer granular-ball while maintaining clear expression of data distribution. To address this, we propose an unsupervised granular-ball partitioning algorithm based on KL divergence. This method evaluates granular-ball quality by calculating the KL divergence between the granular-ball and a Gaussian-distributed control ball, leading to more reasonable divisions. To evaluate the algorithm’s performance, we combine it with the minimum spanning tree-based clustering method and compare it with other unsupervised granular-ball partitioning methods. The experimental results demonstrate that this method effectively balances the efficiency and accuracy of granular-ball partitioning, resulting in more suitable granular-ball representations and improving the effectiveness of the partition.