This small Local AI model could replace an entire document-processing pipeline. ATH-MaaS released OvisOCR3. It’s only a 0.8B-class multimodal model. But one local model can now do all these things 📄 document → Markdown 🧾 document → structured JSON 🔎 OCR + bounding boxes 🧮 formulas → LaTeX 📊 charts → structured data 🔑 key-value extraction 🖼️ figure captioning 🔀 flowcharts → Mermaid 📚 cross-page table reconstruction That last one is interesting. Give it consecutive document pages and OvisOCR3 can determine whether a table continues onto the next page and reconstruct the complete table. Team-reported results OCRBench → 88.8 OCRBench KIE → 90.5 Cross-page table Relation F1 OvisOCR3 → 97.07 Qwen3.8-Max → 94.47 And it’s 🧠 built on Qwen3.5-0.8B 🔓 Apache 2.0 🏠 downloadable/local ⚙️ Transformers + vLLM + SGLang 🔌 OpenAI-compatible local server through vLLM No official GGUF that I can find yet. Link in ALT