OvisOCR2: A 0.8B End-to-End Document Parser Tops Key Benchmarks

Introduction

Document parsing is moving beyond classical OCR pipelines. The OvisOCR2 Technical Report presents a 0.8B model designed to read a document page image and generate a Markdown representation in natural reading order. Instead of only recognizing text, the model is intended to handle a broader set of page elements, including formulas, tables, and visual regions.

This matters because modern document workflows need more than plain text extraction. Research papers, business PDFs, forms, manuals, and knowledge-base inputs often rely on layout, hierarchy, tables, and mathematical notation. A model that can map a page image directly into structured Markdown may reduce the complexity of multi-stage systems and make downstream retrieval, editing, and analysis easier.

Key points

• Compact end-to-end design: OvisOCR2 is a 0.8B document parser that takes a single page image as input and produces Markdown as output.
• Natural reading order: The model aims to reconstruct page content in the order a human would read it, which is especially important for multi-column layouts, mixed text and figures, and complex tables.
• Hybrid data engine: The report describes a data engine that combines filtered real-document annotations with synthetic pages. For synthetic data, rendered images and Markdown targets are derived from the same HTML source, helping align visual input with structured output.
• Multi-stage training recipe: Training includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion.
• Strong benchmark results: On OmniDocBench v1.6, OvisOCR2 reports a state-of-the-art overall score of 96.58. On PureDocBench, it reports the highest Avg3 score of 75.06.
• Additional robustness testing: Beyond public benchmarks, the team also evaluates the model on an internal benchmark covering long-tail and difficult scenarios, where it is reported to achieve the best overall performance among compared methods.

Why it matters

The most notable signal is not just the score, but the shift in architecture. Document parsing leaderboards have often been dominated by pipeline systems that combine layout detection, OCR, table recognition, formula parsing, and post-processing. OvisOCR2 suggests that a unified generative parser can compete at the top of these evaluations while remaining relatively small in model size.

There are still open questions. The available material summarizes the technical report but does not provide independent reproduction results. Real deployments may face harder cases such as noisy scans, very long documents, multilingual layouts, unusual tables, or domain-specific formatting. Even so, OvisOCR2 points to a practical direction: OCR is becoming less about recognizing characters alone and more about converting visual documents into structured, machine-usable representations.

Source: Hugging Face Daily Papers