Tencent released one 5B local model that can parse PDFs, images, charts, AUDIO and VIDEO into the same structured JSON. Instead of stitching together separate OCR, ASR, chart and video pipelines: 📄 PDFs → text, tables, formulas, layout 📊 charts → Markdown 🔀 flowcharts → Mermaid 🎧 audio → speech, speakers, timestamps, sound events 🎥 video → OCR, ASR, actions, camera motion One model → one schema → structured data your software can use in only 5B parameters. Tencent reports 🏆 OmniDocBench v1.6 → 96.96 🥇 OmniParsingBench → 75.08 👉 #1# open-weight On OmniParsingBench it sits behind Gemini-3-Pro overall, but ahead of the other open-weight models in Tencent’s table on Hugging Face Local serving is already supported through: ⚡ vLLM 🚀 SGLang 🤗 Transformers 🔌 OpenAI-compatible local API ⚠️ Official setup currently targets CUDA GPUs. I still don’t see GGUF/llama.cpp support. ⚠️ LICENSE: not Apache 2.0. It uses Tencent’s custom Youtu-Parsing License, which explicitly says: “Youtu-Parsing IS NOT INTENDED FOR USE WITHIN THE EUROPEAN UNION.” 🔗 Link in ALT.