Jina AI Releases jina-ocr-v1 Document Parser for Low-Budget GPUs

Jina AI, a company that is part of Elastic, has launched jina-ocr-v1, an end-to-end visual document parser. This new model is capable of processing various document formats, including PDFs, scans, tables, charts, and invoices, and outputs clean Markdown in a single pass. The model boasts a total of 3.4 billion parameters, with approximately 570 million decoder parameters actively engaged per token. A key feature integrated directly into the checkpoint is a speculative decoding head, which Jina AI developed to enable efficient serving on more accessible GPUs, such as the NVIDIA L4.

In terms of performance, the technical report indicates that jina-ocr-v1 achieved a score of 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench. The model is currently available for research and non-commercial use under an open-weights license. The open weights are approximately 6.8 GB in BF16 format and are compatible with frameworks like Transformers and vLLM. The license governing its use is CC BY-NC 4.0, meaning that any commercial applications require prior contact and agreement with Jina AI.

The jina-ocr-v1 model is built upon DeepSeek-OCR, retaining its two core efficiency components. The DeepEncoder, which comprises about 380 million parameters, integrates SAM (Segment Anything Model), a 16x convolutional compressor, and CLIP-L. This component transforms a 1024x1024 page view, divided into 4,096 patches, into 256 visual tokens. Furthermore, a dynamic-resolution mode can incorporate up to nine additional local tiles, each contributing 100 tokens, allowing a single page to be represented by a maximum of 1,156 visual tokens. The decoder is based on DeepSeek-3B-MoE, featuring 12 layers and a routing mechanism that selects from 64 experts, with 2 experts shared across all operations. The top-6 routing strategy activates roughly 570 million parameters for each token processed. The model supports a substantial position limit of 32,768 tokens. The output format includes Markdown for general text, HTML for tables, and LaTeX for mathematical formulas.

Jina AI incorporated FastMTP Speculative Decoding into jina-ocr-v1, leveraging the near-deterministic and locally structured nature of OCR output. This speculative decoding head, referred to as FastMTP, employs a dense draft block applied recursively for up to K=3 steps. The parameters for these draft blocks remain consistent regardless of the depth of the recursion. The main decoder then verifies these drafts greedily. It accepts the longest sequence of draft tokens that align with its own predictions and then commits one additional token. If all three drafts match the decoder's predictions, an extra token is generated as a bonus. This process ensures that the committed text is equivalent to standard greedy decoding, thereby achieving speedups without any loss in accuracy. With K=3, the model commits an average of 2.73 tokens per step.