Tencent Releases Youtu-Parsing-Omni for Parsing Documents and Media into JSON

On October 9, 2026, the model tencent/Youtu-Parsing-Omni appeared on HuggingFace with the text-generation task. It takes a single input object — a document page, a photograph, a chart, a geometric figure, audio or video — and returns structured JSON with the results of perception and analysis: markup, text, tables, formulas, timecodes, speech recognition and descriptions.

What it means

The model contains 5.3 billion parameters, the largest weights file is model.safetensors at 9.9 GB. It runs via transformers, the weights are published in safetensors format. There are no quantized builds on the site yet, so running it will require a graphics card with enough VRAM headroom for the full weight of the model. License — License Terms of Youtu-Parsing: commercial use is permitted, except in the territory of the European Union. The model may be used, modified and distributed; when distributing, you must include a copy of the license, note the changes and retain the authorship notices. Third-party components are subject to their own licenses. For local running on your own hardware the model is available, but without ready-made quantized builds it is demanding on VRAM. You can follow the appearance of such builds and other news on AI tools in the CPU3D news feed.