MiniCPM-OCR

Converting a scanned document into text is only half the battle when it comes to document automation. Whether table cells are misaligned, whether formula symbols are missed, and whether numbers obscured by shadows in a photo can be directly entered into the database all determine the reliability of subsequent knowledge bases and business systems. What sets MiniCPM-OCR apart is that it provides both a parsing model responsible for recognition and a quality assessment model for evaluating recognition results, giving developers the opportunity to handle “reading” and “determining the level of trustworthiness” separately.

On October 28, 2026, OpenBMB released MiniCPM-OCR 1.0 and MiniCPM-OCR Jev: the former is a document parsing model with 1.3 billion parameters, while the latter is a quality assessment model with 2 billion parameters. The code, model weights, and accompanying documentation are licensed under the Apache-2.0 license, as specified in the project description. It is better suited for developers who need to integrate the system and set up batch processing workflows, rather than as a lightweight scanning tool with just an upload button.Official Release Log - Project Repository

The parsing model is responsible for converting web pages into usable data.

MiniCPM-OCR 1.0 is designed for Chinese, English, and complex document content, covering text, tables, formulas, charts, stamps, and field extraction. Scanning, mobile phone photos, skewed, curved, and re-photographed images are all officially listed as supported scenarios; however, support for a particular input format does not guarantee accurate reproduction for every document.

In practice, full-page parsing involves more than just a single model invocation. The official SDK combines layout detection, region recognition, and reading order organization to generate Markdown; tables are converted to HTML, while equations retain their LaTeX notation. This output is better suited than plain text for further processing—such as search, paragraph segmentation, and structured archiving—and also helps prevent the loss of relationships when concatenating multi-column papers or tables.Modeling and Analysis Process

For example, a team responsible for organizing technical documentation can first parse scanned manuals into Markdown, then create an index by chapter, preserve tables and formulas, and finally feed the data into a retrieval system. If the business ultimately requires invoice amounts or contract dates, the field rules, format validation, and original image locations should also be retained; it is not advisable to simply write the data directly into the business database based solely on a seemingly fluent output.

Jev's value lies in helping to arrange reviews

Jev is not a replacement for OCR recognition. It provides quality scores for candidate parsing results for text, formulas, and table regions; the service can determine whether the original image of the region and the rendered result match by comparing the original image, the candidate content, and the rendered result. It can also score results from other OCR systems, so there is no need to immediately migrate all parsing services just to add a review step.Jev Reasoning Guide

A practical approach is to first prepare a set of business samples with manually verified answers, observe the distribution of errors across different scoring ranges, and then determine which results can automatically proceed to the next stage and which ones need to be queued for review. High-impact fields—such as amounts, ID numbers, and critical dimensions—should undergo stricter checks than ordinary text. Scoring is merely an auxiliary indicator and cannot be directly treated as a guarantee that every field is correct.

When two parsers produce different results for the same section, you can also have Jev score them separately and then select the top candidates. However, the highest score does not necessarily indicate sufficient reliability: if all candidates fall below the business threshold, the task should still be handed over to a human. For content that is obscured, missing, or simply illegible in the original image, quality-based filtering is more useful than forcing the system to generate an answer.

How to Choose Among Common Document Parsing Projects

This table compares product capabilities; it is not a ranking based on standardized testing conditions. Both PaddleOCR and MinerU have comprehensive document processing capabilities; we cannot conclude that MiniCPM-OCR is superior in every recognition task simply because it incorporates quality assessment. For systems already operating stably, it is advisable to first test whether Jev reduces manual screening costs before deciding whether to replace the main parsing pipeline.PaddleOCR Official Repository - MinerU Official Repository

Deployment barriers and costs cannot be assessed based solely on the number of parameters

The project offers an online experience and API integration, as well as an on-premises deployment option. General users can first use public or anonymized samples to observe how the layout is rendered; when dealing with customer contracts, identity documents, and internal financial records, you should first determine upload permissions, the scope of data processing, and the requirements of the selected service. The fact that the open-source code can be deployed in your own environment does not mean that the online demo automatically complies with your organization’s data boundaries.

官方快速开始使用 Linux x86_64 与 NVIDIA 环境，给出的已验证配方为 RTX 5090 32GB、Python 3.12、相应驱动与 CUDA 版本，并要求预留至少40GB磁盘。识别后端与 SDK 需要分别配置。上述是官方验证环境，不是承诺所有用户都必须购买这张显卡，也不能反过来用“1.3B小模型”推导普通笔记本可无门槛处理大批文档。Deployment Guide

Open-source licenses lower the barrier to software adoption, but costs still arise from computing power, storage, operations and maintenance, and manual review. While the repository promotes a free API, specific quotas, duration, and account restrictions are subject to the actual service page and should not be interpreted as unlimited free access. Official performance figures also depend on hardware, page complexity, concurrency, and acceleration settings; when selecting a solution, it is advisable to track the error types, total time per page, and review rate based on your own data.

The most worthwhile scenario for MiniCPM-OCR is the stage in the process of digitizing Chinese scanned documents, technical documentation, and knowledge bases where it has already been recognized that “completed recognition does not mean the text is ready for immediate use.” First, use a small sample to create an error list, integrate parsing with review, and then scale up the processing volume—this approach yields more tangible benefits than simply pursuing a high recognition score on its own.

Official Portal

• Chinese Product Documentation
• Online Experience on the China Site
• Open-Source Code and Model Portal