You give Youtu-Parsing-Omni one file and tell it what kind of thing it is: a document page, a photo, a chart, a flowchart, a geometry figure, an audio clip or a video. It returns one structured file in a fixed layout. For a page that means every block of text, each table, each formula and the reading order, with the position of every element. For a chart it rebuilds the table of numbers behind it. For a flowchart it writes out the boxes and arrows as an editable diagram description. For audio it splits the recording by speaker, transcribes it and stamps the time of each segment. For a lecture video it reads the slides and the speech together and writes a report of the whole thing.
The point is that all of this comes from one 5 billion parameter model with one output format, so a tool built on it does not need a separate text reader, chart reader and transcription service. On the standard document parsing test it edges out every specialised document model in its size class, and across charts, audio and video it trails only Gemini 3 Pro among the models tested.