PepFlow is a multi-modal deep generative model based on the flow-matching framework for designing full-atom peptides that target specific protein receptors. The model characterizes peptide structure using rigid backbone frames in the SE(3) manifold and side-chain angles on high-dimensional tori. It represents discrete residue types as categorical distributions on the probability simplex. By learning the joint distributions of each modality using derived flows and vector fields on corresponding manifolds, PepFlow excels in the fine-grained design of full-atom peptides. The model uses multi-modal paradigm to tackle tasks such as fix-backbone sequence design and side-chain packing through partial sampling. Through experiments, PepFlow demonstrates superior performance in comprehensive benchmarks, highlighting its potential in computational peptide design and analysis. The model is evaluated on three tasks: sequence-structure co-design, fix-backbone sequence design, and side-chain packing. Results show that PepFlow outperforms existing methods in generating peptides with improved binding affinities and structural similarity to native peptides. The model also achieves higher recovery rates and better diversity in sequence design. In side-chain packing, PepFlow outperforms all baselines across all four side-chain angles. The model is capable of generating full-atom peptides with accurate side-chain conformations and dynamics. Despite its success, PepFlow faces limitations in diversity during generation and lacks property-guided generation capabilities. Nevertheless, PepFlow stands out as a potent and versatile tool for computational peptide design and analysis.PepFlow is a multi-modal deep generative model based on the flow-matching framework for designing full-atom peptides that target specific protein receptors. The model characterizes peptide structure using rigid backbone frames in the SE(3) manifold and side-chain angles on high-dimensional tori. It represents discrete residue types as categorical distributions on the probability simplex. By learning the joint distributions of each modality using derived flows and vector fields on corresponding manifolds, PepFlow excels in the fine-grained design of full-atom peptides. The model uses multi-modal paradigm to tackle tasks such as fix-backbone sequence design and side-chain packing through partial sampling. Through experiments, PepFlow demonstrates superior performance in comprehensive benchmarks, highlighting its potential in computational peptide design and analysis. The model is evaluated on three tasks: sequence-structure co-design, fix-backbone sequence design, and side-chain packing. Results show that PepFlow outperforms existing methods in generating peptides with improved binding affinities and structural similarity to native peptides. The model also achieves higher recovery rates and better diversity in sequence design. In side-chain packing, PepFlow outperforms all baselines across all four side-chain angles. The model is capable of generating full-atom peptides with accurate side-chain conformations and dynamics. Despite its success, PepFlow faces limitations in diversity during generation and lacks property-guided generation capabilities. Nevertheless, PepFlow stands out as a potent and versatile tool for computational peptide design and analysis.