This paper introduces a new algorithm for out-of-core mesh partitioning based on a KD-tree. The previous mesh partitioning using KD-tree can produce good results of uniform partitions efficiently. It is still hard to be applied to out-of-core partitioning, in which the whole mesh data cannot be loaded to limited main memory. To build a KD-tree, our algorithm tries to avoid loading and sorting all the vertices of the mesh in the main memory. Our algorithm instead selects minimized set of vertices to sort. A new data structure, a min-heap combined with a max-heap with external files are designed for selective sorting. The size of heaps can be adjusted to each system so our method can be used in systems with limited resources. Test results with huge meshes and comparisons with previous methods, Octree based method and K -means clustering, are shown statistically and visually. A new simplification using our partitioning is also demonstrated to show the utility of our algorithm.