Jina modifies DeepSeek-OCR, ranks first in page throughput among 14 self-tested models.

动察 Beating AI News: Jina AI has released a document parsing model, jina-ocr-v1, which can convert PDFs, scanned documents, tables, and charts directly into Markdown. It is post-trained based on DeepSeek-OCR, retaining its MoE architecture of approximately 3.4 billion total parameters with about 570 million parameters activated per token during decoding, and adds FastMTP speculative decoding.

In Jina's own tests, jina-ocr-v1 scored 91.14 on OmniDocBench v1.6, 0.89 points higher than DeepSeek-OCR-2; it scored 83.4 on olmOCR-Bench, 7.4 points higher than the DeepSeek-OCR base model it is actually based on.

Throughput is one of Jina's main selling points. On a single A100 with a concurrency of 32, it reaches 2.57 pages per second, the highest among the 14 systems tested by Jina, about 22% higher than DeepSeek-OCR's 2.10 pages per second.

The model weights have already been released on Hugging Face under CC BY-NC 4.0, and commercial use requires contacting Jina.