OvisOCR3: a 0.8B OCR model with structured output and cross-page table reconstruction
retained text
OvisOCR3 is out, still based on Qwen3.5-0.8B and available under Apache 2.0, with vLLM inference examples. Compared with OvisOCR2’s Markdown parsing, it adds a few useful options: * Structured JSON output with bounding boxes, element categories, and content, plus a layout-only mode. * Key information extraction using specified fields or a supplied JSON template. * Figure descriptions and Mermaid output for flowcharts. * Cross-page table reconstruction: parse the pages first, then use the same model to merge table fragments across adjacent pages. [OvisOCR3 results](https://preview.redd.it/mo1loqsleswh1.png?width=4336&format=png&auto=webp&s=0789acb753a879e845a245337e1afac65e0c0425) The model card reports **96.74 on OmniDocBench v1.6**, **88.80 on MDPBench**, and **82.31 on PureDocBench**. The bigger reported gains over OvisOCR2 are on MDPBench (69.00 → 88.80) and PureDocBench (75.10 → 82.31). The bounding-box output and cross-page tables seem particularly useful for document workflows. These are the team’s reported numbers, though; I’d be interested in how well the structure holds up on messy scans and long financial tables. Has anyone tried it locally or compared it with GLM-OCR, PaddleOCR-VL, or MinerU? Curious about VRAM usage and processing speed too. HF model card:[https://huggingface.co/ATH-MaaS/OvisOCR3](https://huggingface.co/ATH-MaaS/OvisOCR3) Demo: [https://huggingface.co/spaces/ATH-MaaS/OvisOCR3](https://huggingface.co/spaces/ATH-MaaS/OvisOCR3)
document fetched 2026-10-22 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xbi9h4/ovisocr3_a_08b_ocr_model_with_structured_output