MiniCPM-OCR Jev

Overview

Problem: Document pipelines need a thresholdable signal for whether each OCR region is trustworthy before auto-accepting or sending work to human review.

Built for: Teams parsing scans, photos, or PDFs who want a small open model to score OCR candidates per text, formula, or table region.

First indexed on Jev Directory: 2026-10-27

Creator and team

Name

OpenBMB

Organization

OpenBMB

Handle

@OpenBMB

GitHub

How Jev is used

Role in the product flow

Region-level OCR quality gate: score one candidate, pick best-of-N across parsers, or batch best-of-N across blocks

Primitives

Score

State in

Original region image plus candidate OCR text and category (text, formula, or table). Cloud API, Docker, or SDK paths with a built-in renderer accept image, text, and category; callers without a renderer supply a rendered image of the candidate text.

Decision out

A 0 to 1 quality score when verdict is scored. Precheck, rendering, or forward failures return fixed codes that must not be treated as low model scores.

1. Parse documents and preserve region ids, crops, categories, and candidate text
2. Call score, score_best, or score_best_batch on the API or local vLLM, SGLang, llama.cpp, Ollama, or FlagOS deployment
3. Check verdict equals scored before thresholding
4. Route above-threshold regions to downstream storage; send the rest to human review or another parser

Features and stack

Features

• 2B parameters, Apache-2.0 weights and code
• Scores text, formula, and table regions from 0 to 1
• Free cloud API plus Docker, SDK, vLLM, SGLang, llama.cpp, Ollama, FlagOS
• BF16, GGUF, AWQ, and GPTQ weight formats
• ms-swift fine-tuning guide for domain-specific parsers
• JevOCRBench dataset on Hugging Face for evaluation

Stack

• Qwen3.5-2B derived architecture
• MiniCPM-OCR parsing family
• ModelBest cloud API
• Self-hosted inference (vLLM, SGLang, llama.cpp, Ollama)

Pricing: Apache-2.0 open weights; ModelBest documents a free cloud OCR-Jev API tier. You pay for GPUs or cloud inference you self-host.

Links

• Website
• Repository
• Documentation
• Live demo

FAQ

What does AUROC 93.6 on JevOCRBench mean in practice?

OpenBMB reports AUROC (times 100) of 93.6 on JevOCRBench, meaning the 2B model ranks correct OCR candidates above incorrect ones better than JEV-35B-A3B (90.1) on that vendor benchmark. AUROC does not pick a production threshold; OpenBMB docs say to tune thresholds on your own labeled regions.

When is a score safe to use?

Only when the API or service returns verdict scored. Rendering failures and forward errors return fixed values that must not be read as low confidence from the model.

Is MiniCPM-OCR Jev the same as TypeSafe Jev?

No. It is OpenBMB's independent 2B OCR quality model with Jev in the name and benchmark, not api.typesafe.ai System One. Use TypeSafe Jev listings for hosted choice, score, and noul on general text state.

How do OpenBMB's cross-parser numbers compare?

On 200 JevOCRBench samples scored by MiniCPM-OCR Jev, OpenBMB reports AUROC (times 100) from 91.9 to 93.4 across MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6, and Xiaomi-OCR-0 parsers. Those are OpenBMB-measured on that experiment, not independent audits.