Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs
Back to Home
ai

Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs

September 18, 202656 views2 min read

Jina AI has released jina-ocr-v1, a 3.4B parameter document parser with built-in speculative decoding that enables efficient OCR on low-budget GPUs. The model achieves high accuracy and speed, with open access via Hugging Face and hosted solutions through Jina Reader.

Jina AI has unveiled jina-ocr-v1, a groundbreaking visual document parser designed to transform complex documents like PDFs, scans, tables, and invoices into clean Markdown format. This new model, with 3.4 billion total parameters and approximately 570 million active parameters per token, is built upon the foundation of DeepSeek-OCR and introduces a novel approach to OCR processing. The innovation lies in its integrated FastMTP speculative decoding mechanism, which predicts and generates three tokens per step, significantly enhancing efficiency without sacrificing accuracy.

Performance and Accessibility

The model demonstrates strong performance metrics, scoring 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, highlighting its robustness in handling diverse document types. Notably, it achieves a parsing speed of 2.57 pages per second on a single A100 GPU, making it highly suitable for both high-throughput and resource-constrained environments. This capability is especially valuable for developers and organizations working with limited computational resources, as it enables efficient document processing on low-budget GPUs.

Open Access and Future Implications

Weights for jina-ocr-v1 are available on Hugging Face under the CC BY-NC 4.0 license, promoting open collaboration and research. Additionally, Jina Reader offers hosted access, providing an easy entry point for users who prefer not to manage infrastructure themselves. The release underscores a growing trend in AI development toward democratizing powerful tools, especially in areas like document processing where accessibility and efficiency are paramount. As more models adopt speculative decoding and similar optimizations, we can expect further improvements in speed and usability for OCR applications across industries.

Source: MarkTechPost

Related Articles