Discriminating various leukocyte subsets with specific functions is critical due to their important roles in the development of many diseases. Here, we proposed a general strategy to unravel leukocytes heterogeneity and screen differentiated metabolites as biomarker candidates for leukocyte subtypes using the label-free mass cytometry (CyESI-MS) combined with a homemade data processing workflow. Taking leukemia cells as an example, metabolic fingerprints of single leukemia cells were obtained from 472 HL-60, 416 THP-1, 313 U937, 356 Jurkat, and 366 Ramos cells, with throughput up to 40 cells/min. Five leukemia subtypes were clearly distinguished by unsupervised learning t-SNE analysis of the single-cell metabolic fingerprints. Cell discrimination in the mixed leukemia cell samples was also realized by supervised learning of the single-cell metabolic fingerprints with high recovery and good repetition (98.31 ± 0.24%, -102.35 ± 4.82%). Statistical analysis and metabolite assignment were carried out to screen characteristic metabolites for discrimination and 36 metabolites with significant differences were annotated. Then, differentiated metabolites for pairwise discrimination of five leukemia subtypes were further selected as biomarker candidates. Furthermore, discriminating cultured leukemia cells from human normal leukocytes, separated from fresh human peripheral blood, was performed based on single-cell metabolic fingerprints as well as the proposed biomarker candidates, unveiling the potential of this strategy in clinical research. This work makes efforts to realize high-throughput single-leukocyte metabolic analysis and metabolite-based discrimination of leukocytes. It is expected to be a powerful means for the clinical molecular diagnosis of hematological diseases.