OpenBMB and ModelBest released the 1.3B MiniCPM-OCR 1.0 recognition model and the 2B Jev correctness model. The former parses text, formulas and tables in Chinese and English documents; Jev compares image regions with OCR output and returns a quality score between 0 and 1. It can also assess other OCR engines’ results to prioritize human review. The developers’ reported AUROC of 93.6 on a 100-point scale measures error discrimination, not 93.6% character-recognition accuracy. Code and weights use Apache 2.0. Thresholds need calibration on real document samples, with separate checking for high-impact fields such as contract amounts.