Nanonets' OCR2 Hits 800K Downloads Turning Scanned Docs Into Structured Markdown

Nanonets-OCR2-3B is a compact vision-language model that turns messy documents into clean, LLM-ready markdown with semantic tags for tables, equations, signatures, and more.

·
·
·
Nanonets' OCR2 Hits 800K Downloads Turning Scanned Docs Into Structured MarkdownPRO
  • Nanonets released Nanonets-OCR2-3B, a 3B VLM that converts documents to structured markdown.
  • Built on Qwen2.5-VL-3B-Instruct, Apache 2.0 licensed, trained on 3M+ diverse pages.
  • Handles LaTeX equations, HTML tables, Mermaid flowcharts, signatures, watermarks, and checkboxes.
  • Scores 85.15 on DocVQA and 79.2 on ChartQA, beating Qwen-72B on charts.
  • Runs via transformers, vLLM, SGLang, or GGUF with a 125K context window.
  • Multilingual support across English, Chinese, Arabic, Japanese, and many more languages.

The Nanonets model card reported more than 800,000 downloads at publication. Nanonets-OCR2-3B is a 3B-parameter vision-language model that converts document images into structured Markdown, including semantic labels for signatures, watermarks, equations, checkboxes, tables, and diagrams. Its size, output format, and license make it relevant to developers building local document ingestion and retrieval pipelines. Because Hugging Face counts artifact fetches and repeat downloads, the total cannot establish the number of unique users or production deployments.

The model builds on Qwen2.5-VL-3B-Instruct. Nanonets says it fine-tuned the base model on more than three million pages spanning research papers, financial reports, contracts, healthcare records, tax forms, invoices, and handwritten documents. The release uses Apache 2.0, which permits local use, modification, redistribution, and commercial deployment subject to the license terms.

Page pixels into usable Markdown

Each page enters the model as an image and leaves as serialized content that preserves reading order and document structure. Downstream systems can pass that output to an LLM, index it for retrieval, or parse selected tags into application fields. The advertised schema covers these elements and tasks:

  • Equations: LaTeX with separate inline $...$ and display $$...$$ notation.
  • Tables: Markdown or HTML serialization for layouts that need row and column structure.
  • Form controls: Checkboxes and radio buttons normalized to ☐, ☑, and ☒.
  • Document regions: Headers, footers, page numbers, signatures, and watermarks, including <signature> and <watermark> tags.
  • Diagrams: Flowcharts and organizational charts serialized as Mermaid code.
  • Images: Short descriptions inside <img> tags when captions are absent.
  • Languages and handwriting: Support for handwritten pages and languages including Chinese, Arabic, Japanese, Korean, and Russian.
  • Document questions: Visual question answering with a requested

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Trending
  • No trending articles

Comments

avatar

Next Reads