Identifying peptides directly from data-independent acquisition (DIA) data remains challenging due to the highly multiplexed MS/MS spectra. Spectral library-based peptide detection is sensitive, but it is limited to the depth of the library and mutes the discovery potential of DIA data. We present here, DIA-MS2pep, a library-free framework for comprehensive peptide identification from DIA data. DIA-MS2pep uses a data-driven algorithm for MS/MS spectrum demultiplexing using the fragments data without the need of a precursor. With a large precursor mass tolerance database search, DIA-MS2pep can identify the peptides and their modified forms. We demonstrate the performance of DIA-MS2pep by comparing it to conventional library-free tools in accuracy and sensitivity of peptide identifications using publicly available DIA datasets of varying samples, including HeLa cell lysates, phosphopeptides, plasma, etc. Compared with data-dependent acquisition-based spectral libraries, spectral libraries built directly from DIA data with DIA-MS2pep improve the accuracy and reproducibility of the quantitative proteome.