OvisOCR2 - An end-to-end document parsing model launched by Alibaba's ATH-MaaS team.

What is OvisOCR2?

OvisOCR2 is an end-to-end document parsing model developed by Alibaba's ATH-MaaS team. Trained on Qwen3.5-0.8B and completely open-source, it topped the OmniDocBench benchmark with a score of 96.58. It boasts high-precision recognition and structured extraction capabilities for complex document layouts, supporting the parsing of multi-column layouts, tables, formulas, and other scenarios. It is suitable for intelligent document processing needs such as academic research, archival digitization, and enterprise knowledge base construction.

Main functions of OvisOCR2

• End-to-end document parsingIt can directly output structured text from input document images, without the need for multi-stage pipeline splicing.
• Complex layout recognitionIt supports accurate parsing of complex document layouts such as multi-column layout, mixed text and images, and nested tables.
• High-precision OCR extractionIt achieved first place with a score of 96.58 on OmniDocBench v1.6, leading the industry in character and paragraph recognition accuracy.
• Open source and deployableBased on Qwen3.5-0.8B training, the model weights and code have been open-sourced to HuggingFace.

OvisOCR2 Technical Principles

• Visual Language Model ArchitectureThe model is based on Qwen3.5-0.8B and integrates a visual encoder and a text decoder to achieve end-to-end mapping from images to structured text.
• Large-scale document pre-trainingIt is pre-trained on massive amounts of scanned documents, PDF renderings, and other data to learn the relationship between document layout and semantics.
• End-to-end optimizationIt abandons the traditional multi-stage process of detection → recognition → post-processing, and reduces error accumulation through joint optimization of a single model.
• Lightweight and efficientAchieving state-of-the-art performance with a small model of 0.8B parameters, balancing inference speed and deployment cost.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use OvisOCR2

• Environmental preparationClone the HuggingFace repository and install its dependencies (Transformers, PyTorch, etc.).
• Loading Model:pass AutoModelForCausalLM Load OvisOCR2 weights and corresponding word segmenters.
• Input documentConvert scanned documents or screenshots into tensor format and pass them into the model for forward inference.
• Get ResultsThe model directly outputs structured parsed results in Markdown or plain text format.
• Fine-tuning deploymentBased on our own document data, we continue to fine-tune the layout to adapt to the specific layout requirements of our field.

OvisOCR2's core advantages

• Top of the evaluation listOmniDocBench v1.6 topped the list with a score of 96.58, leading the industry in document parsing accuracy.
• Ultra-lightweightBased on Qwen3.5-0.8B training, state-of-the-art performance can be achieved with only 0.8B parameters, enabling smooth inference on a single card.
• Fully open sourceThe model weights and code have been publicly released on HuggingFace, with no commercial licensing restrictions, and support free secondary development.
• End-to-end architectureThe model abandons the traditional multi-stage pipeline, directly outputting structured text from input document images, significantly reducing engineering maintenance costs.
• Complex layout adaptationIt natively supports complex scenarios such as multi-column layout, mixed text and images, and nested tables, without the need for additional post-processing rules.

OvisOCR2 project address

• HuggingFace model libraryhttps://huggingface.co/ATH-MaaS/OvisOCR2

Comparison of OvisOCR2 with similar competing products

Application scenarios of OvisOCR2

• Digitalization of academic literatureBatch parsing of scanned PDF papers, automatically extracting the main text, figure and table titles, and reference structure, accelerating the organization of research materials.
• Enterprise Records ManagementThis method converts scanned copies of historical paper archives into a searchable structured database, reducing the cost of manual data entry.
• Financial instrument processingIt accurately identifies key fields and nested tables in invoices and contracts, supporting automated financial auditing processes.
• Educational materials organizationIt analyzes the complex multi-column layout of textbooks and test papers, generating structured electronic resources that are easy to search and edit online.
• RAG knowledge base constructionProvides high-quality structured text input for document question-and-answer systems, improving retrieval and enhancing the accuracy of generated answers.