# OpenBMB just put a 1B OCR model on an iPhone, and it can even check its own results — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/5cea5ef7b981261b

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-28 16:31 ingested 2026-10-29 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xhsym7/openbmb_just_put_a_1b_ocr_model_on_an_iphone_and/)

## retained text



```
OpenBMB has posted a video showing MiniCPM-OCR running locally on an iPhone and iPad, parsing real-world documents without sending them to the cloud. 

Handwritten notes, dense formulas, messy tables, low-quality scans. OCR gets a lot harder when the documents aren't perfect.

What's interesting is that this isn't just about extracting text. OpenBMB is also releasing a separate model to check whether the OCR output is actually correct.

* **MiniCPM-OCR 1.0 (1B):** Scores 93.68 on Real-5 OmniDocBench and 54.7 on OCRBench v2.
* **MiniCPM-OCR-Jev (2B):** Evaluates parsed text, formulas, and tables against the original image, assigning correctness scores from 0 to 1. It approaches Qwen3.5-397B-A17B on JevOCRBench and works with outputs from different OCR parsers.

So you get both a lightweight OCR model and a way to catch potentially incorrect results, useful when document accuracy actually matters.

Both have open weights, with support for tools like llama.cpp, vLLM, SGLang, and Ollama.

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

Curious how well the verifier holds up on messy real-world documents. Would you trust a small model to check another model's OCR, or would you still want a human in the loop?
```



document fetched 2026-10-29 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xhsym7/openbmb_just_put_a_1b_ocr_model_on_an_iphone_and
