MiniCPM-V

A high-performance multimodal LLM series designed for efficient deployment on end-side devices like smartphones and laptops.

MiniCPM-V delivers GPT-4V level capabilities within a compact parameter footprint (typically 2B to 8B). Built by OpenBMB and THUNLP, the latest MiniCPM-Llama3-V 2.5 model outperforms larger competitors like Claude 3 Vision on the OCRBench benchmark. It supports 30+ languages, features high-resolution image processing via adaptive tiling, and achieves 150 tokens per second on an iPhone 15 Pro using llama.cpp. This architecture makes advanced visual reasoning (real-time scene description and complex document parsing) viable for local, private hardware.

https://github.com/OpenBMB/MiniCPM-V