# OpenBMB released a 1.3B OCR model and a separate 2B model to check its work — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/e8f1c1e9abc364fd

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-27 17:21 ingested 2026-10-28 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xgyqbb/openbmb_released_a_13b_ocr_model_and_a_separate/)

## retained text



```
A PDF turning into clean Markdown doesn't mean all the numbers survived the trip.

[OpenBMB just released](https://x.com/OpenBMB/status/2115124474882469955) MiniCPM-OCR 1.0, a 1.3B model for parsing Chinese and English documents, alongside MiniCPM-OCR Jev, a 2B model for checking OCR output.

Jev takes an image crop and its corresponding OCR result, then returns a quality score from 0 to 1. It handles text, formulas and tables. It can also score results from other OCR parsers, so using their parser isn't a requirement.

That's the part I'd want in a local RAG pipeline. Feed it a table with a wrong number and your chatbot can very helpfully cite that wrong number back to you. I'd use something like this to flag questionable regions for review before they go into the index.

The code and both models' weights are Apache 2.0. The [GitHub repo](https://github.com/OpenBMB/MiniCPM-OCR) has local deployment instructions and links to quantized weights.
```



document fetched 2026-10-28 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xgyqbb/openbmb_released_a_13b_ocr_model_and_a_separate
