ModelBest and OpenBMB open-source MiniCPM-OCR series with 1.3B parser and 2B quality-check model

ModelBest and OpenBMB jointly announced on October 31 the release and open-source availability of the MiniCPM-OCR series, a family of models for efficient document parsing and parsing-quality assessment. The launch covers two models: MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B).

As the first OCR-focused model in the MiniCPM family, MiniCPM-OCR 1.0 puts high parsing accuracy into a compact 1.3-billion-parameter architecture optimized for edge deployment. According to the announcement, it parses scanned, photographed, curved, tilted and screen-captured documents without cumbersome image pre-correction, lowering the barrier to large-scale document digitization.

On speed, MiniCPM-OCR 1.0 is powered by DSpark decoding acceleration, which the companies say raises end-to-end throughput and cuts per-page latency. For organizations processing millions of pages, such as financial contract archives, academic libraries and government records, that means more documents handled on the same hardware and time budget at materially lower computing cost.

To address page-by-page manual review, the two organizations introduced MiniCPM-OCR Jev, described as the first OCR result decision model and an “AI quality inspector.” It does not perform OCR itself; instead it compares the source image with the OCR output and assigns confidence scores to text, LaTeX formulas and tables, catching missing characters, errors and formatting misalignment. Despite its 2B size, the release claims judgment accuracy typically associated with far larger models.

Jev supports cross-model deployment and can be integrated into existing OCR workflows without replacing the underlying engine. It can route reliable results straight through, flag uncertain output for human review, and pick the best result among multiple OCR systems, moving enterprises from full manual inspection toward risk-based automated quality control.

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.

Hardware coverage is another emphasis. Day-0 adaptation is listed for AMD, Arm, Houmo AI, Intel and NVIDIA, while the FlagOS software stack adds Day-0 support for 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.

The combination pushes OCR from plain recognition toward recognition plus quality control: a small edge-capable model parses, an even smaller model triages and corrects, and decoding acceleration plus broad chip support lower the cost of putting document digitization into production. What to watch next is third-party benchmark performance and real deployment feedback in finance, archives and government document pipelines.