Jina OCR v1
Jina OCR v1 adapts DeepSeek-OCR for document parsing. It reads page images and returns text with document structure, including tables and mathematical content, through weights or hosted Reader access.
How the model works
The vision encoder compresses a 1,024-by-1,024 global image view into 256 visual tokens and adds local tiles. Its mixture-of-experts decoder activates about 570 million parameters per token. Jina adds a shared FastMTP draft head that proposes three tokens ahead. Greedy verification preserves the main decoder's greedy sequence.Jina OCR v1 model details
Use and access
The repository includes custom loading code and examples for Transformers and vLLM. FastMTP acceleration is specific to the documented vLLM route. Jina reports page throughput with the GPU and concurrency stated; those numbers are not general latency promises. Reader can fetch and render a document before calling the model. OCR v1 is the trained parser, while Reader is the surrounding service.Jina OCR v1 model detailsJina OCR v1 technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | jina-ocr-v1jinaai/jina-ocr-v1 |
| Parameters | Approximately 3.4 billion total; 570 million active per tokenJina OCR v1 model details |
| Weight license | CC BY-NC 4.0Jina OCR v1 model details |
| Base model | DeepSeek-OCRJina OCR v1 model details |
Related models
A small vision-language model for questions about images and documents.
Launch date API · Downloadable- ReaderLM v2
Jina AI · ReaderLM
A small language model that turns HTML into Markdown or schema-based JSON.
Launch date API · Downloadable