Description
A skewed scan, a phone photo of an exam paper with shadows, a recaptured newspaper curved into an arc. Traditional OCR hands back misaligned text, broken tables and garbled formulas, and you end up checking every page by hand.
MiniCPM-OCR is a family of open-source document parsing models from OpenBMB. The main one, MiniCPM-OCR 1.0, has only 1.3B parameters and parses Chinese and English documents and PDFs into layout-aware Markdown, covering text, tables, formulas, charts and seals, plus key information extraction. It scores 93.68 on Real5-OmniDocBench, a real-world parsing benchmark, and with DSpark decoding acceleration end-to-end efficiency improves by 55%. The authors report throughput of about 1.49x PaddleOCR-VL-1.6 and 3.62x MinerU2.5 Pro.
The second model, MiniCPM-OCR Jev, is the inspector. Give it an image and its OCR result and it scores each region from 0 to 1, showing which parts to trust and which to review by hand. It can also grade the output of other parsers.
Document parsing: text, tables, formulas, charts and seals, across scans, photographs, curved and skewed pages and recaptured documents.
Key information extraction: pull the fields you need straight out of a document for downstream automation.
Region-level quality scoring: Jev scores text, formula and table regions so human review goes only where it is needed.
Many ways to deploy: vLLM, SGLang, llama.cpp and Ollama from day one, with quantized and GGUF versions, and it can be fine-tuned further.
Easy to integrate: a pip SDK, a full API, free online demos and a SKILL.md that lets an agent get started right away.
Open source: code under Apache-2.0, with weights on Hugging Face and ModelScope.
MiniCPM-OCR is a family of open-source document parsing models from OpenBMB. The main one, MiniCPM-OCR 1.0, has only 1.3B parameters and parses Chinese and English documents and PDFs into layout-aware Markdown, covering text, tables, formulas, charts and seals, plus key information extraction. It scores 93.68 on Real5-OmniDocBench, a real-world parsing benchmark, and with DSpark decoding acceleration end-to-end efficiency improves by 55%. The authors report throughput of about 1.49x PaddleOCR-VL-1.6 and 3.62x MinerU2.5 Pro.
The second model, MiniCPM-OCR Jev, is the inspector. Give it an image and its OCR result and it scores each region from 0 to 1, showing which parts to trust and which to review by hand. It can also grade the output of other parsers.
Features
Document parsing: text, tables, formulas, charts and seals, across scans, photographs, curved and skewed pages and recaptured documents.
Key information extraction: pull the fields you need straight out of a document for downstream automation.
Region-level quality scoring: Jev scores text, formula and table regions so human review goes only where it is needed.
Many ways to deploy: vLLM, SGLang, llama.cpp and Ollama from day one, with quantized and GGUF versions, and it can be fine-tuned further.
Easy to integrate: a pip SDK, a full API, free online demos and a SKILL.md that lets an agent get started right away.
Open source: code under Apache-2.0, with weights on Hugging Face and ModelScope.
