A model with only 2 billion parameters does not read documents: it checks the work of whoever reads them. OpenBMB has released MiniCPM-OCR Jev, a system that, given an image of one document region and a proposed transcription, returns a score between 0 and 1: the higher the score, the more likely the transcription is correct.
The model handles three kinds of regions: text, formulas and tables. In practice, whoever parses the document supplies the original image, the candidate text and the region category, and the system decides whether to accept the result automatically or send it to a person. It can also pick the best among several transcriptions from different OCR models for the same region, and recombine them in reading order.
Image: OpenBMB — https://github.com/OpenBMB/MiniCPM-V/blob/main/docs/minicpm_v4dot5_en.md
Jev was trained from Qwen3.5-2B: the vision encoder and language backbone were kept, while the output head was replaced with a binary classification head that learns to recognise correct transcriptions.
In OpenBMB's reported figures, its discrimination score on JevOCRBench is 93.6 (AUROC times 100), above the 90.1 of JEV-35B-A3B and close to the 93.8 of Qwen3.5-397B-A17B. The size gap is the striking part: the model is far smaller than the competitors cited. On transcriptions produced by MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0, the score stays between 91.9 and 93.4, so the judging ability does not depend on which parser produced the text.
The most useful part for anyone automating digitisation is the threshold. The score is compared with a value the user sets on their own samples: above the threshold a region passes without human review, below it goes to review. Raising the threshold reduces accepted errors but increases manual work, so both the accuracy of accepted regions and the volume of reviews should be monitored. OpenBMB shows charts on legal documents, financial receipts and complex documents illustrating this trade-off.
Jev corrects nothing and creates no new transcriptions: it chooses among those already available. Final quality therefore depends on the coverage and correctness of the candidates. Candidate transcriptions must share the same crop and content boundaries, otherwise the comparison is meaningless.
Image: OpenBMB — https://github.com/OpenBMB/MiniCPM-V/blob/main/docs/minicpm_v4dot5_en.md
The model is available as open weights on Hugging Face in BF16 format, with GGUF variants for llama.cpp and quantised AWQ and GPTQ versions. Guides are provided for vLLM, SGLang, Ollama and FlagOS, plus a fine-tuning path with ms-swift. One practical warning from the documentation: when the service has a built-in renderer, read the score only if the status is "scored"; fixed values returned on errors are not model judgments.
The release comes in the same period as Baidu's Unlimited OCR 2 for long documents and a compact document-analysis model from Xiaomi. The trend is clear: text recognition is shifting toward small models that treat pages as images, while quality control becomes a separate stage of the pipeline.
No comments yet. Be the first!