AI BriefingMiniCPM-OCR-1.0: 1.3B size with 1.49x faster document parsing compared to PaddleOCR
openbmb/MiniCPM-OCR-1.0
About the project
MiniCPM-OCR 1.0 is a 1.3B-scale vision-language model specialized in text-centric document understanding. It supports general text as well as tables, formulas, charts, seals, and structured information extraction, and handles complex documents captured in real-world environments, such as scans, tilted pages, and photos.
It achieves a score of 93.68 on Real5-OmniDocBench, demonstrating top-tier quality among open-source OCR models. Applying DSpark decoding acceleration improves processing efficiency by 55%, providing processing speeds approximately 1.49x faster than PaddleOCR-VL-1.6 and 3.62x faster than MinerU2.5 Pro.
It runs out of the box on major runtimes such as vLLM, SGLang, llama.cpp, and Ollama, and supports quantized variants in BF16 and GGUF formats as well as multi-chip builds. It is provided with a pip SDK, API, and agent workflows for easy integration into development environments, and additional fine-tuning is possible.
It supports English and Chinese documents and is released under the Apache 2.0 license. It is suitable for document automation tasks that require resolving speed bottlenecks in existing OCR pipelines and accurately extracting complex elements such as tables and formulas.
openbmb/MiniCPM-OCR-1.0
The original page has no description.
image-text-to-text
This introduction 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 errors, attribution issues, or removal requests via Contact.