Clustering algorithms based on split-and-merge concept, divisive or agglomerative process are widely developed to extract patterns with different shapes, sizes and densities. Here a multilevel approach is considered in order to characterise general patterns up to finer shapes. This paper focus on the contribution of both fuzzy and evidential models to build a relevant divisive clustering. Algorithms and both a priori and a posteriori split criteria are discussed and evaluated. Basic crisp/fuzzy/evidential algorithms are compared to cluster four datasets within a multilevel approach. Finally, same framework is also applied in embedded spectral space in order to give an overall comparison.