XNews

OpenBMB and ModelBest Pair MiniCPM-OCR Parser with Jev Reliability Scorer

OpenBMB and ModelBest Pair MiniCPM-OCR Parser with Jev Reliability Scorer

OpenBMB and ModelBest released MiniCPM-OCR, a 1.3B-parameter open document parser for Chinese and English scans, photos, curved pages and tilted inputs, paired with a second model, MiniCPM-OCR Jev, that scores the reliability of extracted text, formulas and tables. Jev estimates reliability without fixing errors and also works on outputs from other OCR models, with open weights, a fine-tuning framework and inference support for llama.cpp, vLLM, SGLang and Ollama.

Event version 1.

Key insights

Small parser with broad document coverage

MiniCPM-OCR has only 1.3B parameters. It handles Chinese and English documents, including scans, photos, curved pages and tilted inputs, and supports key information extraction.

Second model scores output reliability

MiniCPM-OCR Jev checks extracted text, formulas and tables. It estimates reliability and does not fix errors itself.

Works on top of other OCR models

Jev also works on outputs from other OCR models, so it can run on top of an existing parser.

Open weights and inference integrations

The weights are open, with a fine-tuning framework and inference support for llama.cpp, vLLM, SGLang and Ollama. Agent Skills integrations are also available.

Suggested pipeline for document agents

A simple setup parses a document with MiniCPM-OCR, scores extracted text, formulas and tables with Jev, and sends low-scoring output to a human or a correction step.

Source posts