XNews

Jina AI releases jina-ocr-v1 document parser

Jina AI releases jina-ocr-v1 document parser

Jina AI released jina-ocr-v1, an end-to-end document parsing model fine-tuned from DeepSeek-OCR, with 3.4B parameters. It achieves 2.57 pages per second on a single L4 GPU, significantly faster than comparable models, and converts full pages into high-quality Markdown.

Key insights

jina-ocr-v1 offers high throughput and accuracy

Jina AI released jina-ocr-v1, a 3.4B parameter document parsing model fine-tuned from DeepSeek-OCR. It processes 2.57 pages per second on a single L4 GPU, several times faster than similar models like chandra-ocr-2 (0.38 pages/s) and dots.mocr (0.55 pages/s). It converts full pages to Markdown with high quality, scoring 83.4 on olmOCR-Bench and 91.14 on OmniDocBench.

Efficient architecture with speculative decoding

The model uses a small vision encoder (380M parameters) and a sparse decoder (DeepSeek-3B MoE) with a FastMTP speculative head, which doubles decoding speed on L4. The architecture reduces visual tokens to 256 per page, enabling faster processing. The model is available via Transformers, vLLM, SGLang, Docker, and Jina's API.

Source posts