OpenBMB releases MiniCPM-OCR 1.0, an open-source 1.3B document parsing model
OpenBMB released MiniCPM-OCR 1.0, a 1.3B open-source OCR model that scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. It supports vLLM, SGLang, llama.cpp, and Ollama, with weights on Hugging Face and ModelScope.
China context
- Original name
- 面壁智能
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate MiniCPM-OCR 1.0 via vLLM, SGLang, llama.cpp, or Ollama, and access a free API for document parsing.
- For investors
- OpenBMB's release of a SOTA open-source OCR model may pressure commercial OCR providers and signal continued investment in efficient, specialized models.
MiniCPM-OCR 1.0 is a 1.3B OCR-specific model for text-centric document understanding, covering text, tables, formulas, charts, seals, and structured information extraction. It achieves SOTA scores on Real5-OmniDocBench (93.68) and OCRBench v2 (54.7), and improves end-to-end processing efficiency by 55% with DSpark decoding acceleration. The model supports Chinese and English documents and PDFs, and is available via Hugging Face, ModelScope, and a free API.
MiniCPM-OCR 1.0 uses DSpark decoding acceleration (DSpark-BF16, DSpark-GGUF) to improve end-to-end processing efficiency by 55%, achieving throughput of approximately 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro. It supports vLLM, SGLang, llama.cpp, and Ollama from day one, with multiple chip platforms and quantized versions available at release.
Developers outside China can now use a high-performance open-source OCR model that outperforms existing open-source alternatives on several benchmarks, reducing the cost and effort of document parsing for applications like RAG and data extraction.
MiniCPM-OCR 1.0 offers a free, open-source alternative to commercial OCR APIs, potentially lowering costs for businesses that need document parsing at scale. Its high throughput and support for multiple deployment frameworks make it suitable for production use.
A measurable signal is whether downstream document processing pipelines integrate MiniCPM-OCR 1.0 and whether independent evaluations reproduce the claimed benchmark scores. The release of MiniCPM-OCR Jev, a region-level evaluation model, suggests future integration for quality control in OCR workflows.