Taxonomic profiling is a challenging first step when analyzing a metagenomic sample. This work presents a method that facilitates fine-scale characterization of the presence, abundance, and evolutionary relatedness of organisms present in a given sample but absent from the training database. We calculate a “k-mer palette” which summarizes the information from all reads, not just those in conserved genes or containing taxon-specific markers. The compositions of palettes are easy to model, allowing rapid inference of community composition. In addition to providing strain-level information where applicable, our approach provides taxonomic profiles that are more accurate than those of competing methods.