> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Tencent releases Youtu-Parsing-Omni, turning documents, audio and video into one JSON output
- URL: https://globalfeed.ai/en/tencent-releases-youtu-parsing-omni-turning-documents-audio-and-video-into-one-json-output/
- Published: 2026-10-10T18:55:21.000Z
- Updated: 2026-10-10T18:55:21.000Z
- Description: Tencent released Youtu-Parsing-Omni, a 5-billion-parameter model that turns documents, photos, charts, flowcharts, geometry figures, audio and video into a single structured JSON output. It tops OmniDocBench v1.6 in the company's table. Its license says the model is not intended for use in the Europ
- Author: GlobalFeed Editor
- Tags: Tencent, Youtu-Parsing-Omni, document parsing, multimodal ai, x-TencentGlobal, dil-en, elle

**Tencent** released a model called **Youtu-Parsing-Omni** on Hugging Face on October 9, 2026\. The model reads many kinds of input and turns the information in them into a single structured JSON output: document pages, photos, charts, flowcharts, geometry figures, audio recordings and video. Models like this are called parsing models; they are used to turn a scanned contract, an invoice or a meeting recording into structured data that software can work with directly. Youtu-Parsing-Omni has about 5 billion parameters and is built on an encoder developed by Tencent's Youtu team and its Youtu-LLM language model.

According to the model card, the output is always a single JSON block. For images it gives the page layout, text, tables, formulas and the position of each element as box coordinates; for audio and video it gives timestamped segments with speech transcripts, on-screen text, sound events and camera motion. Every output also includes a general description of the content. A task prompt selects which kind of output the model produces.

## First on the document benchmark in the company's table

In the table Tencent published, Youtu-Parsing-Omni ranks first on the document parsing benchmark OmniDocBench v1.6 with an overall score of 96.96; in the same table PaddleOCR-VL-1.6 scores 96.34, MinerU2.5-Pro 95.75, Gemini-3-Pro 92.91 and GPT-5.2 86.59\. On OmniParsingBench, which covers documents, charts, geometry, photos, audio and video together, it averages 75.08, second behind Gemini-3-Pro's 77.44 and the highest among open models. Results were also shared for ChemOCR, which reads chemical structures, and PDMX-Synth, which reads sheet music.

## Running it, and the license

Tencent recommends serving the model with its own vLLM plugin, which needs a CUDA GPU, plus ffmpeg for audio and video. The model also runs with Transformers, but that path requires trust\_remote\_code, which executes code shipped with the checkpoint; the model card recommends pinning a reviewed revision. The technical report is in the GitHub repository.

The weights can be downloaded, but the model is released under a custom license called "Youtu-Parsing". The license text states that the model is not intended for use within the European Union and says this clause prevails in case of any conflict. Organizations should review the license terms against their own situation before using it.

All published comparisons were run by Tencent, and the company has not yet released its evaluation code, which is listed as coming soon. For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Hugging Face: Youtu-Parsing-Omni model card](https://huggingface.co/tencent/Youtu-Parsing-Omni?ref=globalfeed.ai), [GitHub: Youtu-Parsing](https://github.com/TencentCloudADP/youtu-parsing?ref=globalfeed.ai), [Youtu-Parsing license text](https://huggingface.co/tencent/Youtu-Parsing-Omni/blob/main/LICENSE?ref=globalfeed.ai).