Jina AI launches jina-ocr-v1: high-performance document parser for low-cost gpus

Jina AI, a division of Elastic, has launched jina-ocr-v1, an end-to-end visual document parser capable of processing PDFs, scans, tables, charts, and invoices into clean Markdown in a single pass. The model, with 3.4B total parameters and 570M active decoder parameters per token, is optimized for low-budget GPUs like the NVIDIA L4. It features speculative decoding and achieves scores of 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench. The model post-trains DeepSeek-OCR, retaining its efficiency components, and includes a FastMTP head for draft generation. It processes up to 1,156 visual tokens per page and outputs Markdown with HTML tables and LaTeX formulas. Jina AI claims it offers the highest throughput among 14 systems, parsing 2.57 pages per second on an A100 40 GB GPU. The model is available for research and non-commercial use under a CC BY-NC 4.0 license, with open weights around 6.8 GB in BF16. Jina AI also provides a hosted API and a reader tool for quick testing, along with resources for self-hosting. The release highlights improvements in throughput and speed, with a 1.95x boost in token per second on an L4 GPU. The model’s performance is benchmarked against other systems like PaddleOCR-VL-1.6 and HunyuanOCR-1.5, though it does not lead in accuracy.