ModelBest and OpenBMB Release MiniCPM-OCR Series for Efficient Document Parsing and Quality Control

ModelBest and OpenBMB have jointly announced the release and open-source availability of the MiniCPM-OCR series, featuring two models designed for document parsing and quality assessment: MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B). According to the announcement, the initiative aims to improve document digitization efficiency and automate quality inspection for enterprise workflows.

Overview of the MiniCPM-OCR Series

The newly released series introduces two primary models targeted at different stages of document processing:

• MiniCPM-OCR 1.0 (1.3B): A compact 1.3-billion-parameter model optimized for edge deployment. It is designed to handle scanned, photographed, curved, tilted, and screen-captured documents without requiring pre-correction.
• MiniCPM-OCR Jev (2B): A 2-billion-parameter decision model that acts as an automated quality inspector, comparing source images against OCR outputs to assign confidence scores and flag errors.

Capabilities and Performance Claims

According to the report, MiniCPM-OCR 1.0 is built to process dense bilingual Chinese-English content, complex mathematical formulas, and multi-level or borderless tables. Beyond full-page reconstruction, the model supports key information extraction (KIE) to pull specified fields from invoices, IDs, and financial statements based on business instructions.

The developers reported that the model uses DSpark decoding acceleration to increase end-to-end throughput and lower per-page latency. Meanwhile, the Jev model is designed to integrate into existing OCR workflows without replacing the underlying OCR engine, automatically identifying reliable results, flagging potential inaccuracies for human review, and selecting optimal results from multiple OCR systems.

Availability and Ecosystem Support

The MiniCPM-OCR series is reportedly available on GitHub, Hugging Face, and ModelScope. The release supports various inference frameworks and tools, including vLLM, SGLang, llama.cpp, and Ollama, alongside SDKs, APIs, Agent Skill plugins, and fine-tuning code. Day-0 hardware adaptation includes support for AMD, Arm, Houmo AI, Intel, and NVIDIA, with broader ecosystem compatibility through the FlagOS software stack.

Verification

Readers interested in verifying the release details, accessing technical documentation, or reviewing evaluation benchmarks can check the official repositories on GitHub, Hugging Face, and ModelScope under the OpenBMB and ModelBest organization profiles.