MiniCPM-OCR Jev

MiniCPM-OCR Jev

Scores how likely an OCR result for a text, formula or table region is correct, from 0 to 1. For teams that auto-accept good results, send the rest to review or choose between OCR models.

Made by openbmb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiniCPM-OCR Jev is a 2B model that judges OCR output instead of producing it. You give it an image of one region and a candidate OCR result, and it returns a score between 0 and 1. A higher score means the candidate is more likely correct.

Its base is Qwen3.5-2B. The usual text output head is swapped for a yes-or-no head that learns to tell right candidates from wrong ones.

Who it suits: It suits teams that run OCR on documents at scale and want to decide which results can pass untouched and which need a person to check.

Industry | Who uses it | How they use it | Example
Legal | Document teams | Screen parsed contract text before review | Send low-score regions to a reviewer
Finance | Back-office teams | Check parsed receipt content automatically | Accept high-score receipt lines
Document AI | OCR engineers | Compare several OCR models on one region | Keep the best-scoring candidate
Research | ML engineers | Study how well a judge model separates right from wrong | Test it on new document types

If you need a model that reads text from images itself, you can skip it, since this one only scores results.

Where to use it: The card describes a cloud API, a Docker image and an SDK service, and you can fetch the files on Windows, macOS or Linux with Python.

• The cloud API
• The Docker image
• The SDK service, with or without a built-in renderer

Use case | Who | What they get back
Automatic screening | Operations teams | A score per region and a threshold for auto-accept
Human review routing | Document teams | A short list of regions that need a person
Best-of-several selection | OCR engineers | The top-scoring result for each region, joined in reading order

How to set it up: You need Python with pip to fetch the files. To run it, use a hosted API, a container image or an SDK service. The weights take about 4.4 GB in BF16.

• Python with pip
• About 4.4 GB of disk space for BF16 weights
• The project's deployment guide for request fields

1. Install the Hugging Face hub package.
2. Download the model files.
3. Run the model through the service you picked.

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-Jev

How to use it: For each region, send the original region image, the candidate text and the region category, which is text, formula or table. You get one score back. Pick a threshold using samples from your own application.

Input | What goes in
Region image | The crop from the document
Candidate text | The OCR result to judge
Category | Text, formula or table

Example: You parse a receipt with an OCR model and send each line region and its text to the scorer. Lines above your threshold pass automatically, and lines below it go to a person.

For newcomers and experts: If you are new, pick a small batch of regions you already know are right or wrong, score them, and see where a sensible threshold falls.

If you know the field, the card measures AUROC on JevOCRBench and reports 93.6. A cross-parser test with four OCR models gave scores between 91.9 and 93.4, so check it on your own data.

OCR model parses a region
Jev scores the candidate
A threshold splits the results
Low scores go to review

Good to know: A higher threshold usually lets fewer results pass automatically, so watch both the accuracy of accepted results and the review load. The card's tests use specific parsers and samples, so validate on your own document types.

• License: apache-2.0
• Cost to run: none named, though the card also mentions a free API
• Data it reaches: region images and OCR text you send, and the model files you download
• Not for: reading text from images on its own, since it only scores candidates

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 2.2B
• Weights: about 4.4 GB in BF16; about 1.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 41,426
• Likes: 96
• License: apache-2.0
• Languages: en, zh
• Based on: Qwen/Qwen3.5-2B
• Quantized versions: 5
• Fine-tunes: 3
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, minicpm-ocr, correctness-decision, ocr-verification, multimodal, vision-language, conversational

Links

• Model: https://huggingface.co/openbmb/MiniCPM-OCR-Jev
• Files: https://huggingface.co/openbmb/MiniCPM-OCR-Jev/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-Jev

Then follow the usage in its model card. Nothing in this kit runs by itself.