MiniCPM-OCR 1.0 and Jev: open OCR model with a suspicious-region verifier

What happened and what to do

The post describes two open models. MiniCPM-OCR 1.0 has 1.3 billion parameters and is built for document reading. MiniCPM-OCR Jev has 2 billion parameters and does not read the document: it scores which regions of the OCR output may contain errors and need human review. The benchmark results cited are presented by the author, with no independent evaluation in the text.

For a company with OCR pipelines feeding RAG or agents, the useful point is the separation between reading and verification. In practice, one can run the reader over documents, pass the output through a verifier that flags low-confidence passages, and send only those passages to human review, tracking flag rate and confirmed error rate. Before adopting it, test both models on a sample of your own documents, with inference on private infrastructure when data is sensitive.

How the consultancy can help

Diagnosis of the current OCR flow and of the points where reading errors reach RAG or agents. Then design and implementation of a pipeline with reading, per-region verification and a human review queue, evaluated on the company's real documents and operated on the client's infrastructure.

Next step

If your documents feed RAG or agents, send a short description of your case (document types, volume and where OCR errors cause problems) and receive a proposal with scope defined by Wendelmaques.