Deepseek OCR - AI Text Extraction

Can I extract chemical formulas and diagrams from scientific PDFs with DeepSeek OCR?

Yes. DeepSeek outputs SMILES strings for chemical structures and includes multimodal capabilities for diagram understanding and figure captions. This makes it suitable for processing scientific documents where formulas and visualizations need to become machine-readable data.

What GPU memory do I need to run DeepSeek OCR on premises?

The base mode requires 8-10 GB of GPU memory. The Gundam mode, which prioritizes maximum accuracy over speed, needs 40 GB. The mode selector lets you trade fidelity for lower resource consumption if you have tighter constraints.

Does DeepSeek OCR preserve table layouts and document formatting?

Yes. It outputs structured formats including HTML and Markdown that preserve layouts, and it is specifically designed to handle tables and diagrams while maintaining their spatial relationships. This makes it useful for forms, invoices, and documents where structure matters for downstream processing.

What does DeepSeek OCR cost per page on the API?

Pricing runs approximately $0.028 per million input tokens, with cached reads charged at a lower rate. At 256 vision tokens per page, costs accumulate quickly at scale, so high-volume users should model their workload before committing. The MIT-licensed open-source version lets you avoid API costs entirely if you operate your own GPU infrastructure.