Youtu-Parsing-Omni

Not yet independently verified. The release date of 9 October 2026 rests only on the creation of the Hugging Face repository; the model card dates its report and weights only as October 2026 and lists the report as coming soon. The benchmark figures are Tencent's own, from the model card, so no score is stored. The licence is a custom one that states the model is not intended for use within the European Union. No independent report was found. We will update this when it can be confirmed, and remove this note.

Youtu-Parsing-Omni is a 5 billion parameter model from Tencent that turns documents, images, charts, geometry figures, audio and video into structured JSON. Its Hugging Face repository was created on 9 October 2026. Its licence states that it is not intended for use within the European Union.

What Youtu-Parsing-Omni does

A single encoder feeds a Youtu-LLM decoder, and the output is one JSON block covering perception (layout, text, tables, formulas, bounding boxes, timestamps, speech recognition and character recognition) and cognition (captions, narratives and reports). The card lists task modes for documents, natural images, charts, flowcharts, geometric figures, audio, natural video and text-rich video. Document output includes LaTeX formulas, tables, Markdown and Mermaid diagrams. Bounding boxes use a 0 to 1000 grid and timestamps use hours, minutes and seconds.

On OmniDocBench v1.6 the card reports an overall score of 96.96, the top of its table. These are Tencent's own results, and the card shows other systems ahead on some tests, for example average similarity on ChemOCR.

Who it is for

The model suits developers and researchers who run their own document and media parsing pipelines and have CUDA GPUs. Serving uses vLLM with an OpenAI-compatible endpoint; audio and video input needs ffmpeg. Loading through Transformers runs code shipped with the checkpoint, so the card advises pinning a reviewed revision.

Key facts

• Maker: Tencent
• Repository created: 9 October 2026
• Licence: the Youtu-Parsing licence. Its first clause says the model is not intended for use within the European Union and that this clause prevails in any conflict
• Size: 5B parameters, BF16 weights
• Pricing: open weights; no price on the model card
• Platforms: Python 3.10 or later with CUDA GPUs
• Weights: Hugging Face, tencent/Youtu-Parsing-Omni