The heterogeneity within a bulk tumor tissue, referred as intratumor heterogeneity, becomes a prevalent confounding factor to tumor genomic studies. Analysis on genomic profilings from heterogeneous tumor samples can potentially lead to false positive differential expression conclusions, and even influence patients’ clinical outcomes and therapeutic responses. To address the intratumor heterogeneity issue, we develop a Fast Tumor Deconvolution (FasTD) tool to separate the pure tumor signals from mixture samples in an efficient way. Assuming a linear combination of the abundance of the mixing components and availability of reference information for the nontumor component(s), our semiparametric regression-based model can quickly provide estimates for the tumor proportion in a mixture, as well as output the tumor specific genomic profile. We demonstrate FasTD is a competitive tumor deconvolution tool for both simulated data and The Cancer Genome Atlas RNAseq datasets, with no requirement for preselected signature genes.