nanonets/Nanonets-OCR2-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Oct 13, 2025Architecture:Transformer0.5K Featherless Exclusive Cold

Nanonets-OCR2-3B by Nanonets is a 3 billion parameter image-to-markdown OCR model designed to transform documents into structured markdown with intelligent content recognition and semantic tagging. It excels at extracting complex elements like LaTeX equations, detailed image descriptions, signatures, watermarks, and tables, making it highly suitable for preparing diverse documents for downstream processing by Large Language Models. The model supports multiple languages and is trained to handle handwritten documents, offering robust performance for various document types.

Loading preview...

Nanonets-OCR2-3B: Advanced Document-to-Markdown Conversion

Nanonets-OCR2-3B is a 3 billion parameter image-to-markdown Optical Character Recognition (OCR) model developed by Nanonets. It specializes in converting diverse documents into structured markdown, focusing on intelligent content recognition and semantic tagging. This model is particularly designed to prepare documents for efficient processing by Large Language Models (LLMs) by providing rich, structured output rather than just raw text.

Key Capabilities

  • Structured Content Extraction: Automatically converts documents into markdown, including semantic tags for various elements.
  • LaTeX Equation Recognition: Identifies and formats mathematical equations into LaTeX syntax, distinguishing between inline and display equations.
  • Intelligent Image Description: Generates structured <img> tags with descriptions for images, charts, and graphs, making visual content LLM-digestible.
  • Signature & Watermark Handling: Accurately detects and isolates signatures within <signature> tags and extracts watermark text into <watermark> tags.
  • Complex Table Extraction: Converts intricate tables into both markdown and HTML formats, preserving structure and content.
  • Flowchart & Organizational Chart Conversion: Extracts these visual elements and represents them as Mermaid code.
  • Multilingual & Handwritten Support: Trained on handwritten documents across multiple languages, including English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, and Arabic.
  • Visual Question Answering (VQA): Can directly answer questions based on document content, responding with "Not mentioned" if the answer is absent.

Good for

  • LLM Pre-processing: Ideal for preparing unstructured document data into a structured, semantically tagged format suitable for LLM ingestion and analysis.
  • Automated Document Processing: Automating the extraction of complex information from legal, business, scientific, and financial documents.
  • Data Digitization: Converting scanned or image-based documents into editable and searchable structured text.
  • Multilingual Applications: Handling documents in various languages and those containing handwritten content.
  • Financial Document Analysis: Specialized prompts and options are available for improved accuracy on table-heavy financial documents.

Finetunes
2 models