MiniCPM OCR

MiniCPM OCR is an open-source document parsing model family that extracts text, tables, and formulas from Chinese and English PDFs and images.

MiniCPM-OCR is an open-source model family designed for efficient document parsing and text recognition. It parses Chinese and English documents and PDFs, handling challenging inputs such as scans, photographs, curved pages, and recaptured documents. The MiniCPM-OCR 1.0 model extracts text, tables, formulas, charts, seals, and structured information into Markdown or JSON formats. The MiniCPM-OCR Jev model evaluates OCR results at the region level, returning a quality score to help users locate content that needs review. The project supports deployment via vLLM, SGLang, llama.cpp, and Ollama, and provides a pip SDK, a free API, and fine-tuning support.

MiniCPM OCR at a glance

Best for: Software engineer, Researcher
Runs on: self-hosted, open source, API
Pricing: Free API
Released: October 2026
Category: AI productivity tool
Website: github.com/openbmb/minicpm-ocr

Key features

Parses Chinese and English documents and PDFs
Extracts text, tables, formulas, charts, and seals
Provides DSpark decoding acceleration for efficient inference
Evaluates OCR results with region-level quality scoring
Supports vLLM, SGLang, llama.cpp, and Ollama backends
Outputs parsed data into Markdown or JSON formats
Includes a pip SDK and full-featured API for integration

Use it for

Parsing scanned or photographed real-world documents
Extracting structured information from complex tables and formulas
Evaluating OCR correctness and identifying content for review
Processing large-scale document collections automatically

Alternatives to MiniCPM OCR

LLM-Aided OCR — It combines Tesseract OCR with local or cloud LLMs to correct errors and format documents.
Docling — Provides advanced PDF understanding and local execution for sensitive data.
RapidOCR — It converts PaddleOCR models to ONNX format for fast inference and deployment.
PaddleOCR — Open-source OCR and document parsing for 100+ languages
docTR — Extracts text and layout details from documents using deep learning.
OpenDataLoader PDF — It ranks first in benchmarks for layout and table extraction without needing a GPU.
PDF Inspector — Detects if a PDF is scanned or text-based to avoid unnecessary OCR costs.
deepdoctection — Orchestrates layout analysis, OCR, and model fine-tuning in PyTorch pipelines.
Chandra — Provides precise document parsing with bounding box data for complex layouts.
MinerU — Parses complex documents and tables into structured markdown for agents.
Jina — It provides reliable reader and search APIs optimized for LLMs and RAG systems.
Sparrow — Extracts structured data locally from documents without cloud dependencies.