OpenBMB releases a small OCR model and a separate AI checker for its output
MiniCPM-OCR pairs a 1.3B document parser with a 2B model that scores whether text, tables and formulas were recognized correctly.
OpenBMB has open-sourced MiniCPM-OCR, a document suite that separates recognition from quality control. MiniCPM-OCR 1.0 is a 1.3-billion-parameter model for Chinese and English documents, while MiniCPM-OCR Jev is a separate two-billion-parameter model that checks OCR output and scores text, tables and formulas.
The parser handles PDFs, scans, photographs, curved pages, charts, seals and structured extraction. It returns Markdown or JSON and supports vLLM, SGLang, llama.cpp and Ollama, along with quantized checkpoints. OpenBMB says its accelerated configuration processes about 1.2 pages per second and can reach higher throughput with concurrent requests.
The verifier can evaluate results from other OCR systems, not only MiniCPM-OCR, and highlights regions that may need review. OpenBMB reports an AUROC score of 93.6 on its JevOCRBench evaluation, close to a much larger Qwen model, though those claims come from the project and need independent reproduction.
For document-heavy agent workflows, this design can be more useful than raw accuracy alone: low-confidence sections can go to a human before incorrect text reaches search, extraction or automation systems. The code, documentation and weights use Apache 2.0.
