OpenBMB releases MiniCPM-OCR 1.0 with a built-in accuracy checker model

OpenBMB has released the MiniCPM-OCR suite, two open-source models aimed at document parsing: MiniCPM-OCR 1.0, a 1.3 billion parameter OCR model, and MiniCPM-OCR Jev, a separate 2 billion parameter model whose job is to check whether the OCR output is actually correct. The idea is that OCR tools can look finished while still getting fields, formulas, or tables wrong, so Jev acts as a model-agnostic checker that can grade text, formulas, and tables from different OCR parsers.

OpenBMB's own benchmark numbers put MiniCPM-OCR 1.0 at 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, and say it leads on OmniDocBench v1.6, Wild-OmniDocBench, and olmOCR-Bench among open-source OCR models compared. MiniCPM-OCR Jev reportedly scores 93.6 on JevOCRBench, ahead of the larger JEV-35B-A3B at 90.1 and close to Qwen3.5-397B-A17B at 93.8.

With a decoding acceleration add-on called DSpark, OpenBMB says end-to-end processing efficiency improves by 55%, and throughput reaches about 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro, with SDK concurrency optimizations pushing throughput up to 128 pages per minute. The model is built on a SigLIP vision encoder paired with a Qwen3.5 language model, supports Chinese and English, and ships with day-one support for vLLM, SGLang, llama.cpp, and Ollama, plus a range of chip vendors through FlagOS.

Why it matters

Pairing a fast OCR model with a dedicated accuracy checker addresses a real gap in document automation, where silent OCR errors can be costly in high-stakes workflows.

Source: OpenBMB on X