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

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 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?