Nanonets-OCR2-3B
Nanonets
Nanonets-OCR2-3B is an open-weights image-to-markdown OCR model fine-tuned from Qwen2.5-VL-3B-Instruct. It turns document images into structured markdown: LaTeX for equations, markdown and HTML for complex tables, mermaid for flow and organisational charts, and dedicated tags for embedded images, signatures, watermarks and page numbers. Checkboxes and radio buttons become Unicode symbols. It is trained on handwritten documents and on documents in English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean and Arabic among others, and answers visual questions about a page, replying "Not mentioned" when the answer is absent.
Strengths
- Converts document images directly into structured markdown intended for downstream LLM processing
- Recognises mathematical equations and emits properly formatted LaTeX, distinguishing inline ($...$) from display ($$...$$)
- Extracts complex tables and returns them in both markdown and HTML table formats
- Extracts flow charts and organisational charts as mermaid code
- Detects and isolates signatures, watermarks and page numbers into dedicated <signature>, <watermark> and <page_number> tags
- Converts form checkboxes and radio buttons into standardised Unicode symbols (checked, unchecked, crossed)
- Trained on handwritten documents and on documents in English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean and Arabic among others
Caveats
- The model card and Hugging Face repository publish no licence at all: despite the weights being public and ungated, no terms for commercial use or redistribution are granted.
- Loses to Gemini 2.5 Flash on ChartQA in the vendor's own VQA table (78.56 vs 84.82).
- The strongest family member, Nanonets-OCR2-Plus, is not released as weights and is reachable only through the hosted Docstrange service; in the vendor's head-to-head table the 3B wins only 29.37% and loses 54.58% against Plus.
- No context window or maximum output length is documented; the max_new_tokens=4096 and max_tokens=15000 values in the card are example arguments in the usage snippets, not model limits.
- Accuracy is input-dependent: the vendor advises raising image resolution, setting repetition_penalty=1 for complex financial tables, and using a different prompt or resolution per document type.
- Vendor
- Nanonets
- License
- Proprietary
- Modalities
- textimage
| Benchmark | Score | Max |
|---|---|---|
| DocVQA (IDP-Leaderboard) | 89.4 | 100.0 |
| ChartQA (IDP-Leaderboard) | 78.6 | 100.0 |
Capabilities
Resources
Pricing
The model card publishes no pricing. The weights are downloadable from Hugging Face and can be run locally with transformers or vLLM. The vendor separately offers hosted access through its Docstrange extraction API (extraction-api.nanonets.com), whose pricing is not stated on the model card.