Jina AI Releases jina-ocr-v1 Based on DeepSeek-OCR

Jina AI has released jina-ocr-v1, a document parsing model built on DeepSeek-OCR for English and Chinese, featuring its own improvements: FastMTP speculative decoding and GRPO post-training with dense verifiable rewards.

Compiled by Sergey KostenchukPublished 2026-09-22Updated 2026-09-22

2026-09-22 Research NVIDIA

🤖 Jina AI Releases jina-ocr-v1 Based on DeepSeek-OCR

The model for parsing scans and PDFs in English and Chinese outputs Markdown with tables and LaTeX formulas: a page is compressed into 256 visual tokens. Jina's contribution is FastMTP speculative decoding and GRPO fine-tuning; on olmOCR-Bench — 83.4 versus 76.0 for the base model.

🌍 Mass translation of archives into Markdown is getting cheaper: 2.57 pages/s on a single A100 — first place in speed among 14 systems, and 91.14 on OmniDocBench v1.6. On NVIDIA L4, FastMTP speeds up decoding almost twofold (≈1.95x).

👤 The weights of jinaai/jina-ocr-v1 are on Hugging Face under CC BY-NC 4.0, for non-commercial use. FastMTP works in vLLM; without installation, you can run URLs through r.jina.ai with a Jina API key. The model also signs images and answers questions about the document.

Source 1: https://jina.ai/news/jina-ocr-v1-faster-document-parsing-on-low-budget-gpus/ Source 2: https://huggingface.co/jinaai/jina-ocr-v1

Sources