LangBoat Intelligence has recently unveiled and open-sourced its cutting-edge MiniCPM-OCR series models. This series encompasses the MiniCPM-OCR 1.0 model, which boasts 1.3 billion parameters, and the MiniCPM-OCR Jev model, featuring 2 billion parameters. Notably, MiniCPM-OCR 1.0 has achieved state-of-the-art (SOTA) performance in two prominent OCR benchmark tests: Real5-OmniDocBench and OCRBench v2. After undergoing DSpark decoding acceleration, its throughput has soared to an impressive 1.217 pages per second.
In the meantime, MiniCPM-OCR Jev stands out as the industry's pioneer OCR parsing result discrimination model, with an Area Under the Receiver Operating Characteristic Curve (AUROC) score as high as 93.6. This model is capable of automatically sifting through recognition results to identify reliable ones and select the optimal result from multiple models. This functionality substantially reduces the workload associated with manual verification.