Facing Wall open-sources two OCR models: 1.3B for recognition, 2B specifically for error detection.

Beating AI News Flash: ModelBest has released and open-sourced the MiniCPM-OCR series. The 1.3B MiniCPM-OCR 1.0 can recognize handwritten text, complex tables, formulas, and skewed or curved documents. It ranks first in four out of six OCR benchmark tests. On an RTX 5090, its processing speed reaches 1.217 pages per second, about 49% faster than PaddleOCR-VL-1.6.

The other model, the 2B MiniCPM-OCR Jev, is specifically designed for error detection. It checks OCR results against the original image and scores text, formulas, and tables separately. High-scoring content can pass directly, while questionable content is sent for manual review. It can also check the output of other models such as GLM-OCR and PaddleOCR without requiring replacement of the existing OCR system.

This design targets the most labor-intensive part of the OCR workflow: machines can read documents quickly, but staff still have to check item by item for missing characters, misaligned lines, or digit recognition errors. ModelBest's internal experiments on 900 manually annotated image blocks show that when the accuracy of the auto-approved portion approaches 100%, up to 85% of legal documents require no manual verification, compared with 56.2% for financial receipts and 35.1% for complex documents.