OpenBMB released MiniCPM-OCR 1.0, a 1.3-billion-parameter open-source OCR model scoring 93.68 on Real5-OmniDocBench for real-world documents such as curved pages, phone photos, tilted scans, and complex layouts. DSpark decoding acceleration improves end-to-end processing efficiency by 55%. A companion model, MiniCPM-OCR Jev (2B), serves as an OCR quality judge to evaluate whether extracted text is trustworthy.

Read original