Jina AI Announces Lightweight Visual Document Parser 'jina-ocr-v1'
Jina AI has released jina-ocr-v1, a visual document parser with 3.4B total parameters (570M active) and built-in speculative decoding. It converts PDFs, scans,
Jina AI has officially announced 'jina-ocr-v1', a new visual document parser, on September 17, 2026. The model features 3.4B total parameters and utilizes a Mixture-of-Experts (MoE) architecture with 570M active parameters to balance performance and efficiency.
Image source: Jina AI (@JinaAI_)
jina-ocr-v1 is optimized to ingest diverse document types—including PDFs, scanned documents, tables, charts, and invoices—and output structured, clean markdown.
Key Features and Technical Details
The model introduces several technical advancements:
- Efficient Architecture: By employing an MoE architecture with 570M active parameters, the model achieves high-quality parsing while maintaining efficient inference performance.
- Speculative Decoding: The model includes built-in speculative decoding mechanisms designed to further accelerate inference speed.
- Broad Document Support: It is engineered to handle complex document structures, particularly extracting data from tables and charts effectively.
How to Use
jina-ocr-v1 is available for developers to integrate into their workflows immediately.
- Hugging Face: The model is published on the Hugging Face Model Hub, allowing for deployment in local or cloud environments.
- Jina Reader API: It can be integrated immediately via the Jina Reader API using the
x-respond-withoption, bypassing the need for manual model deployment.
Due to the MoE architecture, while the active parameter count (570M) is lightweight, the total model size (3.4B) means developers should account for sufficient memory requirements when deploying locally.