Pairs OCR with self-checking, improving reliability of document parsing for agents and RAG.
MiniCPM-OCR is an open document-reading model that also checks whether its own OCR output is trustworthy. OpenBMB released the 1.3B MiniCPM-OCR 1.0, scoring 93.68 on Real5-OmniDocBench for curved and real-world pages, plus an AI judge model.