OpenBMB describes MiniCPM-OCR 1.0 as supporting Chinese and English documents and PDFs, including scans, photographs, curved pages and skewed pages. Its model card lists text, tables, formulas, charts, seals and structured information extraction among its tasks. For whole-document parsing, the SDK first uses PP-DocLayoutV3 to find regions and reading order. MiniCPM-OCR 1.0 recognizes those regions, and the SDK assembles the results. The 1.3-billion-parameter model does not perform that entire pipeline by itself.
Jev’s model card describes a different job: given an image of a text, formula or table region and a candidate OCR result, it returns a score from 0 to 1. A higher score means the candidate is more likely to be correct. The proposed workflow sends lower-scoring regions for human review or compares candidates from different parsers. Jev scores and selects existing results; it does not correct the text itself.
OpenBMB reports a 54.7 score for MiniCPM-OCR 1.0 on OCRBench v2 and throughput about 1.49 times that of PaddleOCR-VL-1.6. These are OpenBMB’s benchmark claims, not independently established performance. The model card describes a throughput test using an NVIDIA GeForce RTX 5090, all 1,651 OmniDocBench v1.6 images, image-level concurrency of one and accelerated configurations for each system. OpenBMB notes that throughput depends on content, backend, concurrency and hardware.
OpenBMB links model weights for both releases and states that the code, documentation and weights use the Apache License 2.0. ModelBest lists MiniCPM-OCR-1.0 as its default document-parsing API model, subject to Console and model-list access. OpenBMB labels its Jev API example “coming soon”, and the Jev model card calls the cloud API guide a stub.
