Open 5B model turns documents, audio and video into one structured format

The Youtu-Parsing-Omni team released a 5-billion-parameter model that converts document pages, images, charts, audio and video into one JSON format, with boxes and timestamps showing where each piece came from, so even a text-only model can read it and cite its source. The authors report 96.96 on the OmniDocBench v1.6 document test, the top score in their comparison, and 75.08 on OmniParsingBench, best among open-weight models but behind Gemini-3-Pro at 77.44. These are the authors' own results; for OmniParsingBench they built their own scoring toolkit, and they say the model is not yet sped up, so it should be slower on dense tables. Code and weights are public. The Hugging Face page has only the authors' note and an automated bot reply, so there is no community discussion yet. Sorted take: worth a small test if you extract data from PDFs, but check speed before you plan around it.