Cell fate transitions emerge from dynamic gene expression programs, yet existing RNA velocity models primarily rely on RNA abundance and globally inferred latent time, limiting their ability to capture local regulatory dynamics. To address these limitations, we introduce MoFlow, a deep neural network that integrates multi-omic data within a relay velocity framework. Unlike previous approaches, MoFlow flexibly infers velocity parameters at single-cell resolution without pre-assigned latent time, enabling a comprehensive and locally adaptive estimation of gene expression kinetics. Applied to single cell multi-omic datasets from brain, skin, and blood cells, MoFlow distinguished chromatin-dependent and independent transcriptional regulation, validated transcription repression models, and identified asynchronous gene repression in neural progenitors. It also uncovered transcriptional activation of DNA damage response genes in radial glia with distinct subnuclear localization. By resolving fine-grained regulatory programs, MoFlow advances the interpretability and precision of RNA velocity analysis beyond the limits of existing models.