OpenBMB open-sources the MiniCPM-OCR suite: a 1.3B parser scoring 93.68 on Real5-OmniDocBench, plus a 2B judge model that flags OCR errors

Background and the problem

Document parsing (turning PDFs, scans and photographed pages into Markdown or JSON) is the entry point for RAG and enterprise agents. Modern OCR models are already accurate, but the hard part is not knowing where they went wrong: one misread number in a contract, receipt or financial statement can be costly, so teams often fall back to reviewing everything by hand. The MiniCPM-OCR suite that OpenBMB (ModelBest) open-sourced on Oct 27 splits the job in two: one model reads, another checks.

How it works

The suite has two models. MiniCPM-OCR 1.0 (1.3B) is the first OCR-specific model in the MiniCPM-V/o family. It handles Chinese and English documents and PDFs, covering text, tables, formulas, charts, seals and key information extraction, and copes with scanned, curved, skewed and recaptured pages; its DSpark decoding heads improve end-to-end efficiency by 55%, according to the team. MiniCPM-OCR Jev (2B) is an OCR quality judge: given an image and any parser's output, it returns a 0-to-1 quality score per region, so developers can set a threshold and route low-scoring regions to human review. Code, docs and weights are all Apache 2.0.

Benchmarks (vendor-reported)

MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, both claimed as state of the art. Throughput is about 1.2 pages per second, roughly 1.49x PaddleOCR-VL-1.6 and 3.62x MinerU2.5 Pro, and up to about 128 pages per minute with SDK concurrency. Jev reaches 93.6 AUROC on JevOCRBench, above JEV-35B-A3B at 90.1 and close to Qwen3.5-397B-A17B at 93.8, and stays between 91.9 and 93.4 on outputs from GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0. No independent third-party results yet; the repo includes a reproduction guide.

Getting started

Day-one support covers vLLM, SGLang, llama.cpp and Ollama, with GGUF, AWQ and GPTQ builds. On hardware, Arm, Intel (OpenVINO), NVIDIA and Houmo AI are supported at launch, plus 10 more chip vendors through FlagOS. The quickest path is to have a coding agent follow the repo's SKILL.md to set up the environment, download weights and parse a first document; a pip SDK and the ModelBest platform API are also available. Teams with an existing OCR pipeline can add Jev purely as a quality check without replacing their parser. Fine-tuning is supported through LLaMA-Factory and Swift.