OpenBMB/MiniCPM-OCR

OpenBMB/MiniCPM-OCR

Description: An open-source OCR solution for fast, accurate document parsing with correctness assessment.

Language: Python

License: Apache-2.0

Stars: 55

Forks: 3

Open issues: 0

Created: 2026-10-27T13:22:00Z

Pushed: 2026-10-28T02:37:16Z

Default branch: main

Fork: no

Archived: no

README:

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.

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 document parsing: MiniCPM-OCR 1.0 achieves a SOTA score of 93.68 on Real5-OmniDocBench, a benchmark for real-world document parsing, and demonstrates leading parsing capabilities on OmniDocBench v1.6, Wild-OmniDocBench, and olmOCR-Bench. It supports Chinese and English documents and PDFs, handling both standard documents and challenging capture conditions, including scans, curved pages, photographs, skewed pages, and recaptured documents.
• ⚡ Efficient inference. With DSpark decoding acceleration (**DSpark-BF16**, **DSpark-GGUF**), MiniCPM-OCR 1.0 improves end-to-end processing efficiency by 55%. Its throughput is approximately 1.49× that of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro, reducing the time needed for large-scale document parsing.
• 🔍 General recognition and information extraction: MiniCPM-OCR 1.0 achieves a SOTA score of 54.7 on OCRBench v2, demonstrating strong general OCR capabilities. It also supports key information extraction (KIE), extracting relevant document content for downstream information management and automated processing.
• 🛠️ Easy deployment and development: MiniCPM-OCR 1.0 supports vLLM, SGLang, llama.cpp, and Ollama from day one, with multiple chip platforms and quantized versions available at release. It supports further fine-tuning and provides a [pip SDK](#local-deployment) and a [full-featured API](#api) for application integration and customization. You can also [get started quickly with SKILL.md](#skill). Try the Online Demo (CN / Global) to explore parsing results.

SOTA Performance, Higher Throughput

Among the open-source OCR models shown below, MiniCPM-OCR 1.0 achieves a parsing throughput of 1.2 pages/s and the highest equally weighted average score across four document parsing benchmarks: OmniDocBench v1.6, Real5-OmniDocBench, Wild-OmniDocBench, and olmOCR-Bench.

The average score weights OmniDocBench v1.6, Real5-OmniDocBench, Wild-OmniDocBench, and olmOCR-Bench equally. The horizontal axis shows throughput in pages per second on a logarithmic scale. See the [reproduction guide](./docs/benchmark_reproduction_en.md) for test settings and reproduction details, and the [model guide throughput evaluation](./docs/minicpm_ocr_1dot0_en.md#throughput) for each model's throughput.

Benchmark Comparisons

View complete MiniCPM-OCR 1.0 evaluation results

See the [OCR model guide](./docs/minicpm_ocr_1dot0_en.md#evaluation) for per-benchmark bar charts and throughput details.

Examples

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

English Documents

Chinese Documents

More Examples

Original documents and parsed results, side by side.

English Documents

Chinese Documents

MiniCPM-OCR Jev

MiniCPM-OCR Jev is an OCR quality 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.

• 🛡️ Strong performance in a small model: With only 2B parameters, MiniCPM-OCR Jev achieves an AUROC (×100) of 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8).
• 🧩 Consistent generalization: Across parsing results from MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6, and Xiaomi-OCR-0, its AUROC (×100) ranges from 91.9 to 93.4.

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.

Review-Free Accuracy

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

Decision Examples

Jev scores for text, formula, and table recognition.

See the [model guide](./docs/minicpm_ocr_jev_en.md), [use cases](./docs/minicpm_ocr_jev_en.md#use-cases), [cross-model evaluation](./docs/minicpm_ocr_jev_en.md#validated-parsers), and [decision examples](./docs/minicpm_ocr_jev_en.md#examples).

Using MiniCPM-OCR

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

Skill

Ask a coding agent to follow the MiniCPM-OCR 1.0 [skill](./docs/SKILL_en.md) 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

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

API

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

[MiniCPM-OCR 1.0 API example](./docs/minicpm_ocr_api_en.md)  ·  [MiniCPM-OCR Jev API example](./docs/minicpm_ocr_jev_api_en.md) (coming soon)

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](./docs/minicpmvocr_tutorial_en.md) · [MiniCPM-OCR...