Jiangsu Suning Bank uses DeepSeek-VL2 to read credit files, with recognition accuracy above 97%

The problem

Banks must read large piles of paperwork before they lend. Credit applications arrive as scans, photos and tables nested inside other tables. The bank says its earlier text-recognition software, known as OCR, struggled with low table recognition rates, handwriting and photographed documents.

What they built

Jiangsu Suning Bank, called 苏商银行 in the source, applies DeepSeek-VL2 to credit materials. DeepSeek-VL2 is a multimodal model, meaning it handles text and images together. It is released by DeepSeek as an open-weight model, so its files can be run and adapted by the bank itself. The model reads nested tables and image files and extracts their content.

How they did it

The article gives few technical details. The bank describes combining the multimodal model with a mixture-of-experts design, an approach where different parts of a model handle different kinds of input. It does not state the team, cost or timeline. The article says banks in general are cautious at the start, picking a few scenarios first and widening use later.

Results

A bank representative says overall recognition accuracy for credit materials rose to over 97%. The same person says efficiency of the whole credit review process improved by 20%. These are the bank's own statements, relayed by the newspaper. No before-and-after data, sample or measurement period was published.