ModelBest and OpenBMB open-source MiniCPM-OCR for parsing and quality control

MiniCPM-OCR 1.0: compact parsing for real-world documents

According to the announcement, MiniCPM-OCR 1.0 is the first OCR-focused model in the MiniCPM family. It has 1.3 billion parameters and an architecture optimized for edge deployment. The companies say it pushes forward the balance between efficiency and performance for open-source OCR models.

The model is built for real-world conditions. It parses scanned, photographed, curved, tilted and screen-captured documents without cumbersome image pre-correction, which the companies say lowers the barrier to large-scale document digitization. It handles dense bilingual Chinese-English content, complex mathematical formulas, and multi-level or borderless tables. According to ModelBest and OpenBMB, it performs at a leading level across a broad range of mainstream industry evaluations.

Beyond reconstructing full pages, it supports key information extraction (KIE). This means it can pull specified fields from invoices, IDs and financial statements according to business instructions.

The model uses DSpark decoding acceleration, which the companies say significantly increases end-to-end throughput while lowering per-page latency. For enterprises processing millions of pages, such as financial contract archives, academic libraries and government records, this means handling far more documents with the same hardware and time budget, at a substantially lower computing cost, according to the release.

MiniCPM-OCR Jev: an AI quality inspector

The second model, MiniCPM-OCR Jev, targets the long-standing need for page-by-page manual review. It is described as the first OCR result decision model. Jev does not perform OCR itself. Instead, it compares the source image with the OCR output and quickly assigns confidence scores to text, LaTeX formulas and tables. It detects missing characters, errors and formatting misalignments.

Despite its 2B-parameter size, Jev delivers judgment accuracy that the companies say is typically associated with far larger models. It supports cross-model deployment and can be added to existing OCR workflows without replacing the underlying OCR engine.

• It can identify reliable results for direct processing.
• It can flag potentially inaccurate outputs for human review.
• It can select the best result from multiple OCR systems.

According to the companies, this lets enterprises move from full manual inspection toward risk-based, automated quality control.

Availability and hardware support

The MiniCPM-OCR series is available on GitHub, Hugging Face and ModelScope. It supports vLLM, SGLang, llama.cpp and Ollama, and comes with SDKs, APIs, Agent Skill plugins and open-source fine-tuning code.

Day-0 hardware adaptation covers AMD, Arm, Houmo AI, Intel and NVIDIA. Through the FlagOS software stack, the series also has 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 (Arm).

What it means for document-heavy organizations

For organizations that digitize large archives, the release combines two pieces: a small parsing model aimed at edge deployment and a separate checker that can sit on top of any OCR engine. Because Jev works across models, teams that already run an OCR system could add automated review without switching engines. The open-source release, the broad runtime support and the fine-tuning code also give developers room to adapt the models to their own workflows. The performance claims come from ModelBest and OpenBMB, and the announcement does not include specific benchmark figures.