OvisOCR2: Compact 0.8B End-to-End Model for Document Parsing

LinkATH-MaaSTool

ATH-MaaS has released OvisOCR2, a lightweight 0.8-billion parameter multimodal model designed for page-level document parsing. Built on Qwen3.5-0.8B, it converts document page images directly into Markdown containing tables, LaTeX formulas, and visual bounding boxes.

Key points

  1. End-to-end page parsing: Converts document images into natural-reading-order Markdown, rendering formulas as LaTeX, tables as HTML, and images with bounding-box tags.
  2. Architecture and training: Post-trained on top of Qwen3.5-0.8B using a dataset combining real-world and synthetic data via a multi-stage recipe of SFT, RL, and OPD.
  3. Benchmark performance: Achieves an overall score of 96.58 on OmniDocBench v1.6, topping the leaderboard as the first end-to-end model over pipeline methods; scores 75.06 Avg3 on PureDocBench.
  4. Deployment and licensing: Fully supported via Transformers, vLLM, and SGLang under an Apache-2.0 open-source licence.

Why it matters

Traditional document OCR and extraction setups rely on multi-step pipelines that are complex to deploy and maintain. OvisOCR2 demonstrates that a sub-billion-parameter end-to-end model can outperform traditional pipelines while maintaining a very small operational footprint.

Analysis

By producing Markdown directly annotated with bounding boxes, HTML tables, and LaTeX math, OvisOCR2 streamlines ingestion pipelines for RAG systems and digital document archives. While the small footprint enables cost-effective self-hosting, the authors note that complex real-world layouts may still cause repetitions or missing content, prompting the inclusion of truncation and cleanup logic in the sample inference implementation.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Video

    Video:Fahd Mirza

    Testing Alibaba Ovis-OCR2: Compact 0.8B Model for Document Parsing

    Fahd Mirza tests Alibaba Ovis-OCR2, an open-source 0.8B document parsing model that outputs clean Markdown. Despite tiny hardware requirements, the end-to-end model outperforms traditional pipeline methods on major OCR benchmarks.

    AI & LLMs· Demo

  • Link:docling.ai

    Docling: Open-Source Document Intelligence for AI and RAG

    Docling is an open-source tool hosted by LF AI & Data that converts complex documents like PDFs, Office files, and images into structured data for AI applications. It operates locally and fully offline, preserves layouts and table structures, and provides interfaces for Python, CLI, REST APIs, and MCP.

    AI & LLMs· Tool

  • Repository:docling-project/docling

    Docling: Document Parsing for Generative AI

    Docling is an open-source Python tool designed to parse documents across multiple formats—such as PDF, office files, audio, and HTML—to prepare them for generative AI workflows.

    66kstarsPython

    AI & LLMs· Tool

  • Link:Apodex AI

    Apodex-1.1-mini: Multimodal Agentic Model Released on Hugging Face

    Apodex AI has released Apodex-1.1-mini under the Apache-2.0 licence on Hugging Face. Based on the qwen3_5_moe architecture, the model is designed for long-horizon research and agentic workflows with native tool execution.

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.