OpenBMB, the open-source group behind the MiniCPM model family, has released MiniCPM-OCR, a pair of small open models for reading documents and for checking what was read. The GitHub repository went public on October 27, 2026, with weights on Hugging Face, and OpenBMB and its partner company ModelBest announced the series in a press release on October 31. Code, documentation and the weights of both models are published under the Apache 2.0 licence.
The first model, MiniCPM-OCR 1.0, has 1.3 billion parameters and is the first OCR-specific model in the MiniCPM-V family. It parses Chinese and English documents and PDFs into Markdown or JSON, covering running text, tables, formulas, charts, seals and key-information extraction, and is meant to cope with scans, photographs, curved or skewed pages and documents photographed from a screen. OpenBMB reports a score of 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, which it describes as state of the art among open models, and a throughput of about 1.2 pages per second. It says its DSpark decoding acceleration improves end-to-end efficiency by 55% and puts throughput at about 1.49 times that of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro.
The second model is the more unusual part of the release. MiniCPM-OCR Jev, with 2 billion parameters, does not perform recognition itself. Given an image region and a candidate transcription of text, a formula or a table, it returns a quality score between 0 and 1, so that reliable regions can pass straight through and doubtful ones can go to a person. It is designed to work with other parsers too: OpenBMB reports an AUROC of 93.6 on its own JevOCRBench, against 90.1 for a 35-billion-parameter judge and 93.8 for Qwen3.5-397B-A17B, and values between 91.9 and 93.4 when judging output from MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0.
The documentation is careful about what that score means. It tells users not to treat the published results as a ready-made confidence threshold, and to label a sample of their own documents, measure accuracy and review volume at different cut-offs, and only then pick a setting. AUROC measures how well a scorer separates right from wrong answers, not how accurate the accepted text is.
Deployment options are broad from the start. The models run on vLLM, SGLang, llama.cpp and Ollama, with BF16, GGUF, AWQ and GPTQ weights, a pip SDK, a free hosted API on ModelBest's platform and fine-tuning recipes for LLaMA-Factory and Swift. The repository also ships SKILL.md files that let a coding agent install the models and parse a first document on its own. The press release lists day-zero support for chips from AMD, Arm, Intel and NVIDIA as well as Huawei Ascend and several other Chinese accelerator makers.
By November 2 the repository had about 760 stars and the main model page roughly 400 likes. AI.info and Eigen Radar covered the release, relying on OpenBMB's material; no independent benchmark has been published. The launch comes days after LlamaIndex released its own 4B parsing model, OCRia-1, in a crowded field of compact document models.
For teams building retrieval and document pipelines, the notable idea is the separate judge: a small open model that estimates where OCR is likely wrong could cut manual checking, provided its thresholds are tuned on the team's own documents.