OvisOCR2 - An End-to-End Document Parsing Model Launched by Alibaba ATH-MaaS Team

What Is OvisOCR2

OvisOCR2 is an end-to-end document parsing model launched by the Alibaba ATH-MaaS team, trained based on Qwen3.5-0.8B and fully open source. The model topped the OmniDocBench v1.6 evaluation with a score of 96.58, possesses high-precision recognition and structured extraction capabilities for complex layout documents, supports parsing of scenarios such as multi-column layouts, tables, and formulas, and is suitable for document intelligence processing needs such as academic research, archive digitization, and enterprise knowledge base construction.

OvisOCR2's Main Features

• End-to-End Document Parsing: Input a document image and directly output structured text, without the need for multi-stage pipeline stitching.
• Complex Layout Recognition: Supports precise parsing of complex document layouts such as multi-column layouts, mixed text and image layouts, and nested tables.
• High-Precision OCR Extraction: Ranked first on OmniDocBench v1.6 with a score of 96.58, with character and paragraph recognition accuracy leading the industry.
• Open Source and Deployable: Trained based on Qwen3.5-0.8B, with model weights and code already open-sourced on HuggingFace.

OvisOCR2's Technical Principles

• Vision-Language Model Architecture: The model uses Qwen3.5-0.8B as its base, integrating a vision encoder and text decoder to achieve end-to-end mapping from images to structured text.
• Large-Scale Document Pretraining: Pretrained on massive data such as scanned documents and PDF renderings to learn document layout and semantic associations.
• End-to-End Optimization: Abandons the traditional multi-stage process of detection -> recognition -> post-processing, and reduces error accumulation through joint optimization by a single model.
• Lightweight and Efficient: Achieves SOTA performance based on a small model with 0.8B parameters, balancing inference speed and deployment cost.

Follow WeChat and reply "Open Source" to join theAI Open Source Project Communication Group

How to Use OvisOCR2

• Environment Preparation: Clone the HuggingFace repository and install dependency libraries (Transformers, PyTorch, etc.).
• Load the Model: Load the OvisOCR2 weights and corresponding tokenizer via AutoModelForCausalLM .
• Input the Document: Convert the scanned file or screenshot into tensor format and pass it to the model for forward inference.
• Get the Results: The model directly outputs structured parsing results in Markdown or plain text format.
• Fine-Tuning and Deployment: Continue fine-tuning based on your own document data to adapt to layout requirements in specific domains.

OvisOCR2's Core Advantages

• Top of the Evaluation: Topped OmniDocBench v1.6 with a score of 96.58, with document parsing accuracy leading the industry.
• Extremely Lightweight: Trained based on Qwen3.5-0.8B, achieving SOTA performance with only 0.8B parameters, with smooth inference on a single GPU.
• Fully Open Source: Model weights and code have both been made public on HuggingFace, with no commercial licensing restrictions and support for free secondary development.
• End-to-End Architecture: The model abandons the traditional multi-stage pipeline and directly outputs structured text from input document images, significantly reducing engineering maintenance costs.
• Complex Layout Adaptation: Native support for complex scenarios such as multi-column layouts, mixed text and image layouts, and nested tables, without the need for additional rule-based post-processing.

OvisOCR2 project address

• HuggingFace model repository：https://huggingface.co/ATH-MaaS/OvisOCR2
• arXiv technical paper：https://arxiv.org/pdf/2607.13639

Comparison of OvisOCR2 with similar competing products

Application scenarios of OvisOCR2

• Digitization of academic literature: Batch parse scanned PDF papers, automatically extract the main text, figure and table captions, and reference structure, accelerating the organization of research materials.
• Enterprise archive management: Convert scanned copies of historical paper archives into a searchable structured database, reducing manual data entry costs.
• Financial document processing: Accurately recognize key fields and nested tables in invoices and contracts, supporting automated financial review processes.
• Educational material organization: Parse the complex multi-column layouts of textbooks and exam papers, generating structured electronic resources for easy online retrieval and editing.
• RAG knowledge base construction: Provide high-quality structured text input for document question-answering systems, improving the answer accuracy of retrieval-augmented generation.