On this page
OpenBMB’s MiniCPM-V README News marks [2026.10.28] for open-sourcing the MiniCPM-OCR suite—a pair of models for document parsing / OCR plus result checking, not a general MiniCPM-V multimodal release and not Baidu’s Unlimited OCR line. Product materials and Hugging Face cards license the suite under Apache License 2.0, with weights on Hugging Face, code on GitHub, day-0 support for vLLM, SGLang, llama.cpp, and Ollama, plus pip SDK, API, and online demo links on the product page. Related OCR reading from another vendor: Baidu Open-Sources Unlimited OCR 2 for Long Documents.
Two roles: parser and checker
MiniCPM-OCR 1.0 is the parser. OpenBMB lists it at about 1.3B parameters and says it is trained on MiniCPM-V 4.6. It targets Chinese and English documents and PDFs under hard capture conditions—scans, phone photos, curved or tilted pages, and re-photographed screens. Product materials also list automatic upright orientation correction and adaptive upsize for dense text and fine digits, plus an LLaVA-UHD v4 tiled vision path for high-resolution pages. This package is an OCR-specific suite; it is not the MiniCPM-V 4.7 on-device multimodal release. Same-family multimodal reading: MiniCPM-V 4.7: Open 35B Multimodal Model That Runs on Phones.
MiniCPM-OCR Jev is about 2B parameters and does not parse document text. Given a region image and a candidate OCR string, it scores correctness / quality (about 0–1). OpenBMB frames it as a model-agnostic checker that can sit after different parsers for routing, human review, or best-of-N selection. The product page calls it a System-1 style gate: one forward prefill for the score, no autoregressive decode. The product page covers text, formula, and table candidates.
Vendor-reported scores and throughput
OpenBMB’s product page and model cards report, for MiniCPM-OCR 1.0, Real5-OmniDocBench 93.68 and OCRBench v2 54.7, and list further OmniDocBench / Wild-OmniDocBench / olmOCR-Bench figures in the same vendor tables. For Jev they report JevOCRBench 93.6 (AUROC×100), alongside their own comparisons to much larger models on that bench (for example they cite Qwen3.5-397B-A17B at 93.8 and JEV-35B-A3B at 90.1 in the same vendor table). With DSpark speculative decoding, they claim about 55% end-to-end efficiency gain, roughly 1.2 pages/second, about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro under their stated test setup. All of those numbers are vendor-reported—for reference only.
Where to get it
Weights live on Hugging Face as openbmb/MiniCPM-OCR-1.0 and openbmb/MiniCPM-OCR-Jev. Code and SDK notes live in the MiniCPM-OCR GitHub repo; the product page also points to pip SDK, API, and regional / global playgrounds. Related open-weights reading: Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face.
Availability, demos, quantization builds, and serving recipes follow OpenBMB’s product page, Hugging Face cards, and GitHub README. For reference only.