MiniCPM-OCR README

README проекта с github.com/OpenBMB/MiniCPM-OCR. Одним файлом markdown: README.md.

MiniCPM-OCR: fast, accurate document parsing into text, tables, and formulas, with a lightning arrow connecting documents to structured results; a decision model helps identify regions that need review (illustration).

MiniCPM logo

MiniCPM-OCR 1.0 Guide Hugging Face ModelScope   MiniCPM-OCR Jev Guide Hugging Face ModelScope

Demo (CN / Global)  ·  Get Started  ·  Weights  ·  Free API  ·  Skill  ·  Local Deployment  |  中文
Join the Community via Feishu / Lark  ·  Discord

MiniCPM-OCR is a family of models for efficient document parsing and text recognition. MiniCPM-OCR 1.0 parses Chinese and English documents and PDFs, balancing quality and efficiency across challenging inputs such as scans, photographs, curved pages, and recaptured documents. MiniCPM-OCR Jev evaluates OCR results at the region level, distinguishing problematic regions from reliable ones. It works with different parsing models to help users locate content that needs review.

📢 News

MiniCPM-OCR 1.0

MiniCPM-OCR 1.0 (1.3B) is the first OCR-specific model in the MiniCPM-V/o family. It focuses on text-centric document understanding, covering text, tables, formulas, charts, seals, and structured information extraction. Its key features include:

SOTA Performance, Higher Throughput

Among the open-source OCR models shown below, MiniCPM-OCR 1.0 achieves 1.2 pages/s and the highest overall performance on OmniDocBench v1.6, Real5-OmniDocBench, Wild-OmniDocBench, and olmOCR-Bench.

Throughput and parsing quality of open-source OCR models. The horizontal axis shows throughput in pages per second on a logarithmic scale; the vertical axis shows the equally weighted average score across four document parsing benchmarks.

See the reproduction guide for test settings and reproduction details, and the model guide throughput evaluation for each model's throughput.

Benchmark Comparisons

Radar chart comparing MiniCPM-OCR 1.0 with other OCR models across six benchmarks; some values remain to be confirmed.

View complete MiniCPM-OCR 1.0 evaluation results

Detailed OCR model results on OmniDocBench v1.6, Real5, Wild, olmOCR-Bench, CHAOS-Bench, and OCRBench v2.

See the MiniCPM-OCR 1.0 guide for per-benchmark bar charts and throughput details.

Examples

MiniCPM-OCR 1.0 and Jev promo video. Click to watch the demo on YouTube.

Left to right: original document, detected layout, and rendered Markdown.

English Documents

English document animation: original, detected layout, and rendered Markdown.

More English Examples

Original documents and parsed results, side by side.

English examples comparing inputs and results: handwriting, historical and scanned documents, newspapers, financial statements, and key information extraction.

Chinese Documents

Chinese document animation: original, detected layout, and rendered Markdown.

More Chinese Examples

Original documents and parsed results, side by side.

Chinese examples comparing inputs and results: handwritten notes, complex formulas, newspapers, financial statements, stamped documents, illustrations, and key information extraction.

MiniCPM-OCR Jev

MiniCPM-OCR Jev is an OCR correctness decision model for text, formulas, and tables. Given an image and its corresponding OCR result, it returns a quality score from 0 to 1 to help select reliable results and locate content that needs review.

Evaluation Results

The table below reports overall and per-category AUROC (×100) on JevOCRBench. Higher values indicate better discrimination between correct and incorrect parsing results.

Overall, text, formula, and table AUROC results on JevOCRBench, highlighting MiniCPM-OCR Jev.

Review-Free Accuracy

Accuracy at different review-free rates across three document types.

Accuracy versus review-free rate for legal documents, financial receipts, and complex documents.

Set the threshold only after checking it on a small batch from the real application.

Decision Examples

Jev scores for text, formula, and table recognition.

Original images, two OCR candidates, and their Jev scores for text, formulas, and tables.

See the MiniCPM-OCR Jev model guide for more on usage, technical specifications, evaluation results, and efficiency results.

Using MiniCPM-OCR

Choose how to integrate MiniCPM-OCR 1.0 and MiniCPM-OCR Jev:

How Document Parsing Works

Full-page parsing runs as a two-stage pipeline orchestrated by the minicpm-ocr SDK. The two stages use separate models, run in separate environments, and may run on different machines:

flowchart LR
    DOC["PDF / image"] --> S1

    subgraph SDKENV ["minicpm-ocr SDK (its own environment)"]
        S1["Stage 1 · Layout detection<br>PP-DocLayoutV3 layout model (PaddlePaddle / PaddleX)<br>regions, reading order, cropping"]
        ASM["Assemble output<br>Markdown / JSON in reading order"]
    end

    subgraph SRVENV ["Recognition service (separate environment)"]
        S2["Stage 2 · Region recognition<br>MiniCPM-OCR 1.0<br>vLLM / SGLang / llama.cpp / Ollama"]
    end

    S1 -- "each cropped region, over HTTP,<br>with a prompt for its region type" --> S2
    S2 -- "text, tables, formulas, charts, seals" --> ASM

A full local deployment therefore prepares two environments side by side: the SDK environment (pip package minicpm-ocr, plus PaddlePaddle / PaddleX for the layout model) and the inference-service environment of the chosen backend. The Skill walks through the complete setup — backend choice, environment installation, model downloads, and verification — and the SDK tutorial covers the same steps for manual deployment. For region prompts and per-stage details, see the document parsing workflow.

Skill

Ask a coding agent to follow the MiniCPM-OCR 1.0 skill to set up the environment, download models, and parse your first document. Send the following instruction to the agent:

Read and execute https://github.com/OpenBMB/MiniCPM-OCR/blob/main/docs/SKILL_en.md

Ask a coding agent to follow the MiniCPM-OCR Jev skill to deploy, start, and verify the scoring service locally — either from the Docker image or with the SDK, using vLLM, SGLang, llama.cpp, or Ollama as the inference backend. Send the following instruction to the agent:

Read and execute https://github.com/OpenBMB/MiniCPM-OCR/blob/main/docs/SKILL_jev_en.md

If you have already cloned this repository, you can use ./docs/SKILL_en.md and ./docs/SKILL_jev_en.md directly.

Docker Deployment

Deploy MiniCPM-OCR 1.0 with the official Docker image, which bundles model weights, the SDK, vLLM, and the document parsing backend. The Docker Hub page provides startup instructions for both the OCR model API and the end-to-end document parsing API, supporting image and PDF parsing into Markdown / JSON.

Requires Linux x86-64, an NVIDIA GPU, a driver compatible with CUDA 12.9, and Docker configured with NVIDIA Container Toolkit. Follow the image documentation to select and start the service you need.

Free API

Use the Free API to access MiniCPM-OCR 1.0 recognition and MiniCPM-OCR Jev quality scoring.

MiniCPM-OCR 1.0 API docs  ·  MiniCPM-OCR Jev API docs

MiniCPM-V Demo — iOS, Android & HarmonyOS

MiniCPM-OCR 1.0 parses documents entirely on your device, even offline. This document recognition feature is currently available on iOS only.

On-device demo tutorial

Local Deployment

Run the models on your own hardware: MiniCPM-OCR 1.0 parses PDFs and images into Markdown or JSON, while MiniCPM-OCR Jev scores the parsing results.

MiniCPM-OCR 1.0 deployment tutorial · MiniCPM-OCR Jev deployment tutorial · Web demo deployment tutorial

Fine-Tuning

Model Fine-tuning tutorial
MiniCPM-OCR 1.0 LLaMA-Factory · Swift
MiniCPM-OCR Jev Swift

Model Weights

Model Description Main Model Weights DSpark Heads
MiniCPM‑OCR 1.0 Efficient document parsing and text recognition for Chinese and English documents and PDFs, balancing quality and speed across scans, photographs, curved pages, and recaptured documents. 🤗  ModelScope BF16
🤗  ModelScope GGUF
🤗  ModelScope AWQ
🤗  ModelScope GPTQ
🤗  ModelScope BF16
🤗  ModelScope GGUF
MiniCPM‑OCR Jev An OCR correctness decision model for text, formulas, and tables. Given an image and its OCR result, it returns a quality score from 0 to 1. 🤗  ModelScope BF16 🤗  ModelScope GGUF
🤗  ModelScope AWQ 🤗  ModelScope GPTQ

💡 FAQ and Troubleshooting

See the FAQ for common questions about installation, model downloads, GPU memory, input formats, and unexpected outputs.

📄 License

The entire project is released under the Apache License 2.0, including code, documentation, and model weights for MiniCPM-OCR 1.0 and MiniCPM-OCR Jev.

MiniCPM-OCR (OpenBMB/MiniCPM-OCR)

Ищем в базе…

Что это такое?MiniCPM-OCR — это пара моделей для понимания документов. MiniCPM-OCR 1.0 разбирает китайские и английские документы и PDF-файлы в структурированный текст, включая таблицы, формулы, диаграммы и печати. MiniCPM-OCR Jev — отдельная модель, которая оценивает результаты такого разбора по шкале от 0 до 1 и отмечает участки, требующие ручной проверки.
Это одна модель или два отдельных продукта?Две отдельные модели, которые работают вместе. MiniCPM-OCR 1.0 выполняет собственно разбор и распознавание документов, а MiniCPM-OCR Jev — самостоятельная модель проверки корректности, которая проверяет результат первой модели по каждому участку.
Как начать пользоваться?Есть три варианта: напрямую вызвать бесплатный API для распознавания и оценки качества, попросить ИИ-агента для программирования выполнить инструкции из прилагаемых файлов SKILL.md для автоматической настройки окружения и загрузки моделей или развернуть модели локально на своём оборудовании.
Какие бэкенды для инференса нужны, чтобы его запустить?Он работает на vLLM, SGLang, llama.cpp или Ollama без дополнительной настройки, поддерживает квантованные веса (BF16, GGUF, AWQ, GPTQ), а также имеет SDK, устанавливаемый через pip, и полный API для интеграции в ваши собственные приложения.
Могу ли я сразу доверять оценке качества Jev?Не слепо. В README рекомендуется сначала подобрать порог принятия и проверки на выборке из ваших собственных документов, потому что точность при разной доле документов, пропускаемых без проверки, зависит от типа документа.
Насколько большие модели?У MiniCPM-OCR 1.0 1,3 млрд параметров, у MiniCPM-OCR Jev — 2 млрд. Это заметно меньше, чем у многих альтернатив в классе OCR/VLM, при этом модели показывают результаты на уровне лучших современных решений.
Насколько быстрее других OCR-инструментов?С ускорением декодирования DSpark пропускная способность MiniCPM-OCR 1.0 примерно в 1,49 раза выше, чем у PaddleOCR-VL-1.6, и в 3,62 раза выше, чем у MinerU2.5 Pro. Скорость составляет около 1,2 страницы в секунду.
Могу ли я дообучить эти модели на своих данных?Да, для обеих моделей есть отдельные руководства по дообучению с использованием LLaMA-Factory и Swift.
Что можно извлечь помимо обычного текста?Распознаются таблицы, формулы, диаграммы и печати. Также поддерживается извлечение ключевой информации (KIE) для получения структурированных данных из документов.
Какие есть похожие проекты?Xiaomi-OCR-0, surya, GLM-OCR, MiniCPM-V, decider, Valen, GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold, olmocr
Автор
OpenBMB
Размер
65.7 MB
Звёзды
977
Возраст
22 д
Коммиты · 90 дней