← sources

A 1.3B OCR model just hit 93.68 on real-world document parsing and there's now an AI judge for OCR errors

source: reddit (reddit:LocalLLaMA)published 2026-10-27 16:21ingested 2026-10-28 02:38

retained text

Most OCR models focus on reading documents. **MiniCPM-OCR adds a second piece: a model that checks whether the OCR result is actually trustworthy.**

OpenBMB just released two open-source models:

* **MiniCPM-OCR 1.0 (1.3B):** Scores **93.68 on Real5-OmniDocBench**, targeting real-world documents like curved pages, phone photos, tilted scans, and complex layouts. DSpark decoding acceleration boosts end-to-end processing efficiency by 55%.
* **MiniCPM-OCR Jev (2B):** An OCR quality judge that compares the original image with candidate text, catching errors in formulas, tables, and text. It achieves **0.936 AUROC**, close to a 397B-parameter model on the same evaluation.

The interesting part is Jev can also evaluate outputs from other OCR systems, so you don't necessarily have to replace your existing pipeline.

Both models are open source, with support for vLLM, SGLang, llama.cpp, and Ollama.

Hugging Face Links:   
[https://huggingface.co/openbmb/MiniCPM-OCR-Jev](https://huggingface.co/openbmb/MiniCPM-OCR-Jev)  
[https://huggingface.co/openbmb/MiniCPM-OCR-1.0](https://huggingface.co/openbmb/MiniCPM-OCR-1.0)

document fetched 2026-10-28 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xgx26j/a_13b_ocr_model_just_hit_9368_on_realworld