Chromatin immunoprecipitation followed by deep sequencing (ChIP-Seq) became a method of choice to locate DNA segments bound by different regulatory proteins. ChIP-Seq produces extremely valuable information to study transcriptional regulation. The wet-lab workflow is often supported by downstream computational analysis including construction of models of nucleotide sequences of transcription factor binding sites in DNA, which can be used to detect binding sites in ChIP-Seq data at a single base pair resolution. The most popular TFBS model is represented by positional weight matrix (PWM) with statistically independent positional weights of nucleotides in different columns; such PWMs are constructed from a gapless multiple local alignment of sequences containing experimentally identified TFBSs. Modern high-throughput techniques, including ChIP-Seq, provide enough data for careful training of advanced models containing more parameters than PWM. Yet, many suggested multiparametric models often provide only incremental improvement of TFBS recognition quality comparing to traditional PWMs trained on ChIP-Seq data. We present a novel computational tool, diChIPMunk, that constructs TFBS models as optimal dinucleotide PWMs, thus accounting for correlations between nucleotides neighboring in input sequences. diChIPMunk utilizes many advantages of ChIPMunk, its ancestor algorithm, accounting for ChIP-Seq base coverage profiles ("peak shape") and using the effective subsampling-based core procedure which allows processing of large datasets. We demonstrate that diPWMs constructed by diChIPMunk outperform traditional PWMs constructed by ChIPMunk from the same ChIP-Seq data. Software website: http://autosome.ru/dichipmunk/