Scanned pages, handwritten notes and complex tables can be converted into structured Markdown in a single pass by jina-ocr-v1, a new OCR model from Elastic. It is designed to replace the multi-stage pipelines that older OCR relies on for complex layouts.
The model is a vision-language model trained for one task. It outputs the text in an image in the order a human would read it, preserving its structure, across more than 100 languages.
Conventional OCR on complex documents typically chains together page segmentation, element classification, text recognition and reassembly. Elastic says each step can introduce errors that accumulate, leading agents and RAG pipelines to return answers that don't reflect the source.
The model retains headings, sections, lists and reading order in Markdown output. Tables are converted into basic HTML suitable for import into spreadsheets or other applications.
It reads handwriting, including block text in many languages and English cursive. Printed mathematical formulas are converted into LaTeX.
The model also recognises elements that should be excluded, such as page numbers, boilerplate headers and footers, and text that appears incidentally inside images.
Input can include images of varying quality from source formats such as PDF, PPTX and XLSX. The model does not natively accept PDFs, so pages must first be converted to an image format such as PNG.
jina-ocr-v1 is based on the DeepSeek-OCR architecture, with FastMTP added to accelerate inference through multi-token prediction.
Elastic says it scores 83.4 on the olmOCR-bench benchmark, the highest published score among models with fewer than 600 million active parameters. The model is roughly one tenth the size of the benchmark leader.
On OmniDocBench, it scores 91.14 overall. According to Elastic, that is higher than GPT-5.2 and the largest Qwen3 vision-language model, and close to Gemini 3.
"Traditional OCR pipelines break down with complex layouts, tables, handwriting and other highly visual content," said Han Xiao, vice president of AI, Elastic. "Until now, companies either had to accept those limitations or pay a significant premium to use general-purpose LLMs for ingesting documents."
On OmniDocBench measures of character-level accuracy and reading order, Elastic says jina-ocr-v1 outperforms all three frontier LLMs it tested. Only one model, Ovis2.6-30B-A3B, scored higher in all OmniDocBench areas, with about six times as many active parameters.
Images are resized to 1024x1024, with up to nine additional higher-resolution tiles for detail. Elastic notes that features too small to make out at 2048x2048 will probably not be read correctly.
Availability
jina-ocr-v1 is available now through the Elastic Inference Service, included with Elastic Cloud, with GPU acceleration and no need to provision infrastructure.
It can also be accessed pay-per-token through the Jina API, or run on-premises using pre-built containers under a commercial licence from Elastic. The model is available on Hugging Face under CC BY-NC 4.0 for academic and non-commercial use.
A detailed technical blog includes benchmark comparisons and sample outputs.