The Pulse
OpenBMB Pairs a 1.3B OCR Model With a 2B Error Judge
OpenBMB released MiniCPM-OCR 1.0, a 1.3-billion-parameter model for reading Chinese and English documents, alongside MiniCPM-OCR Jev, a separate 2-billion-parameter model that scores OCR results. The company reports benchmark results for bo
AI.info Team · Written with AI · fact-checked ·
OpenBMB released MiniCPM-OCR 1.0 and its companion model, MiniCPM-OCR Jev, on October 27, pairing document recognition with a separate model that judges whether the extracted text is reliable. The 1.3-billion-parameter parser reads Chinese and English documents; the 2-billion-parameter Jev model scores text, formula and table results against the corresponding image region. Both are available as open-source components under Apache License 2.0, according to the project repository.
A parser and a separate quality check
MiniCPM-OCR 1.0 handles text, tables, formulas, charts, seals and key-information extraction. The broader document workflow also uses PP-DocLayoutV3 to locate regions and reading order, then sends cropped regions to the recognition model and assembles the output into Markdown. OpenBMB says the system handles PDFs as well as scans, photographs, skewed pages, curved pages and recaptured documents.
Jev does not run OCR itself. It takes a document-region image and a candidate result, then returns a score from 0 to 1; higher scores mean the model judges the candidate more likely to be correct. Teams can route low-scoring regions to human review or compare candidates produced by different OCR systems. OpenBMB says Jev can assess results from other parsers, including GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0.
What OpenBMB’s benchmark figures measure
OpenBMB reports a score of 93.68 for MiniCPM-OCR 1.0 on Real5-OmniDocBench and 54.7 on OCRBench v2. The company also reports throughput of about 1.2 pages per second across four document-parsing benchmarks and says its DSpark decoding method improves end-to-end processing efficiency by 55%. Those are vendor-reported results; the repository links to a reproduction guide for test settings.
For Jev, OpenBMB reports an AUROC of 93.6 on JevOCRBench, compared with 90.1 for JEV-35B-A3B and 93.8 for Qwen3.5-397B-A17B. AUROC measures how well a scoring system separates correct from incorrect candidates; it is not the same as the accuracy of accepted OCR text. In a separate cross-parser test using 200 JevOCRBench samples, the company reports AUROC values from 91.9 to 93.4 across four parsers.
The review threshold still belongs to the user
OpenBMB’s Jev documentation cautions against treating an AUROC result as a ready-made confidence threshold. It recommends collecting and labeling examples from the intended application, measuring both the accuracy of automatically accepted regions and the review volume at different thresholds, then checking the chosen setting on independent samples. The docs say the published results do not establish a universal threshold.
That qualification matters for workflows involving small figures, dense tables or unfamiliar document formats: a quality score can help prioritize checking, but it does not guarantee a correct extraction. OpenBMB also says Jev selects among existing candidates, so its output remains limited by the coverage and correctness of the OCR systems being compared.
Open weights, several deployment routes
The repository lists vLLM, SGLang, llama.cpp and Ollama among supported runtimes, and offers model weights in BF16, GGUF, AWQ and GPTQ formats. OpenBMB’s parser guide describes the full-document pipeline, from PDF or image input through layout detection and region recognition to Markdown output. The release gives developers a small parser and a separate quality scorer; deciding what score is safe enough to accept still requires testing on their own documents.