Two models released
ModelBest and OpenBMB have released and open-sourced the MiniCPM-OCR series, which comprises two models: MiniCPM-OCR 1.0 at 1.3 billion parameters and MiniCPM-OCR Jev at 2 billion. The companies said in the announcement that MiniCPM-OCR 1.0 is the first OCR-focused model in the MiniCPM family and that it is built for deployment on edge devices. The 1.3B model uses DSpark decoding acceleration, which the companies said raises end-to-end throughput and lowers latency per page.
Parsing without pre-correction
The companies said the model parses scanned, photographed, curved, tilted and screen-captured documents without image pre-correction. It handles dense Chinese-English content, mathematical formulas and multi-level or borderless tables, and it can pull specified fields from invoices, IDs and financial statements according to business instructions, a task the companies call key information extraction.
A check on the output
MiniCPM-OCR Jev does not perform OCR itself. The companies describe it as an AI quality inspector that compares the source image with the OCR output and assigns confidence scores to text, LaTeX formulas and tables, detecting missing characters, errors and formatting misalignments. They said it can be added to existing OCR workflows without replacing the underlying OCR engine, that it can select the best result from multiple OCR systems, and that it can pass reliable results straight through while flagging uncertain output for human review.
Where it is available
The series is available on GitHub, Hugging Face and ModelScope, and supports vLLM, SGLang, llama.cpp and Ollama, along with SDKs, APIs, Agent Skill plugins and open-source fine-tuning code, the companies said. Hardware support at launch covers AMD, Arm, Houmo AI, Intel and NVIDIA, and through the FlagOS software stack it extends to AI and edge chips from CIX Technology, Huawei Ascend, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, Sunrise, T-Head, TsingMicro, NVIDIA and Apple Mac on Arm.
