Introducing MiniCPM-OCR: High-Performance AI-Powered Document Parsing by ModelBest and OpenBMB

ModelBest and OpenBMB Debut MiniCPM-OCR Series

On October 31, 2026, ModelBest and OpenBMB made an exciting announcement regarding the release of the MiniCPM-OCR series, a cutting-edge open-source suite of models designed for efficient document parsing and quality assessment. The launch features two models: MiniCPM-OCR 1.0, which consists of 1.3 billion parameters, and MiniCPM-OCR Jev, equipped with 2 billion parameters.

The MiniCPM-OCR 1.0 Model

MiniCPM-OCR 1.0 is the first model to focus exclusively on optical character recognition (OCR) within the MiniCPM family. It stands out with its impressive parsing accuracy, all within a streamlined architecture tailored for edge deployment. This model has been specifically developed to excel in real-world conditions, adeptly handling various document types, including scanned files, photographs, curved documents, and even content captured on-screen.

The model can operate without the need for extensive image pre-corrections, making it incredibly valuable for organizations looking to digitize vast quantities of documents. Furthermore, it supports bilingual Chinese-English content, complex mathematical formulas, and various table formats, making it highly versatile. Through its full-page reconstruction capabilities, key information extraction (KIE) features enable businesses to generate specific outputs from invoices, identification documents, and financial statements based on given criteria.

With the addition of DSpark decoding acceleration technology, MiniCPM-OCR 1.0 significantly boosts end-to-end processing speeds while reducing latency for each page processed. For enterprises that deal with large volumes of documents, such as financial institutions and academic libraries, this means they can achieve greater output levels without increased hardware or operational costs.

Introducing MiniCPM-OCR Jev

In conjunction with the MiniCPM-OCR 1.0 launch, ModelBest and OpenBMB also introduced MiniCPM-OCR Jev, which serves as a pioneering model for evaluating OCR output quality. Unlike standard OCR models, Jev focuses on analyzing the results produced by other OCR engines, providing confidence scores for the extracted text and data. By detecting errors, formatting inconsistencies, and missing information, Jev performs a critical function in ensuring the integrity of OCR outputs.

Despite its more compact size, Jev's judgment capabilities rival those of much larger models, thanks to its advanced algorithms. The model can be seamlessly integrated into existing OCR workflows, promoting a move toward automated quality control processes. It helps organizations reduce the need for time-intensive manual reviews by automatically flagging potentially inaccurate results for human intervention, thus streamlining the overall productivity of document processing tasks.

Availability and Integration

The MiniCPM-OCR series has been made available on popular platforms, including GitHub, Hugging Face, and ModelScope. It supports various configurations, including vLLM, SGLang, llama.cpp, and Ollama, complemented by SDKs, APIs, and fine-tuning codes that are open-source. The system also boasts compatibility with a range of hardware, including AMD, Arm, Intel, and NVIDIA.

By proactively adopting this innovative technology, organizations can harness the power of AI-driven document processing and management systems to improve efficiency and maintain high-quality output without excessive costs.

Whether transforming outdated paper archives into digital collections or automating the processing of incoming documentation, the MiniCPM-OCR series stands as a testament to advancements in AI. With energy efficiency and improved performance at the core, these tools empower organizations across various sectors to engage with the digital future confidently.

In summary, with the launch of the MiniCPM-OCR series, ModelBest and OpenBMB are redefining the standards for document parsing and quality control, establishing a new benchmark in the field of AI technology for businesses.