AIAsia

MiniCPM-OCR pairs document parsing with a separate quality checker

ModelBest and OpenBMB released the suite on October 28. A compact parser and a distinct verification model address different stages of document digitization.

iNewsChannel IntelligenceAI-assisted reporting · Source-linked analysis

Source published: Oct 28, 2026
Published: Nov 3, 2026

1 min · Reading time

OpenBMB chart comparing OCR throughput with average scores across four document benchmarks — ModelBest / OpenBMB contributors · Apache License 2.0 · Original developer-reported comparison: throughput on the horizontal logarithmic axis and an equal-weight average of four benchmarks vertically. Model-card timing uses an RTX 5090 and image-level concurrency 1; results depend on configuration.

Added after our earlier coverage. The original source date is shown below; this is not a new announcement today.

What happened

MiniCPM-OCR 1.0 is a 1.3-billion-parameter model for Chinese and English documents, including scans and photographed pages. MiniCPM-OCR Jev is a separate 2-billion-parameter model that compares source images with OCR output and scores text, formulas and tables.

The official project licenses code, documentation and both models’ weights under Apache 2.0. Downloadable checkpoints, deployment guides, demos and an API route are linked from the repository. The original changelog dates the launch October 28, before the October 31 international press release.

iNC / Analysis

Why it matters

Our analysis: separating extraction from verification could help organizations direct human review toward uncertain passages. A confidence score is not proof of correctness, however. The supplied comparison chart reports the developers’ tests; it does not establish equivalent results on a publisher’s own documents.

What to watch

Watch for independent testing on unfamiliar layouts and languages, and whether verification catches costly omissions rather than simply agreeing with the parser.