OpenBMB Releases MiniCPM-OCR 1.0 and Jev for Document Parsing and Quality Evaluation

·2026.10.28 01:21

Key point

The 2B-parameter MiniCPM-OCR Jev achieves 0.936 AUROC, comparable to the performance of a 397B model for detecting OCR errors.

OpenBMB released two open-source models: MiniCPM-OCR 1.0 (1.3B) for document parsing and MiniCPM-OCR Jev (2B) for quality evaluation.
MiniCPM-OCR 1.0 achieves a score of 93.68 on Real5-OmniDocBench, handling complex layouts like curved pages and tilted scans.
DSpark decoding acceleration improves end-to-end processing efficiency by 55%.
MiniCPM-OCR Jev reaches 0.936 AUROC, comparable to a 397B-parameter model, and can evaluate outputs from other OCR systems without replacing existing pipelines.
Both models support vLLM, SGLang, llama.cpp, and Ollama.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.