MiniCPM-V

MiniCPM-V is a line of open-weight vision-language models from OpenBMB, a research lab spun out of Tsinghua University, built to run vision and video understanding on phones and other resource-limited devices rather than requiring a server GPU. The May 2024 release, MiniCPM-Llama3-V 2.5, pairs an 8B parameter Llama3 backbone with a SigLip vision encoder and claims performance on OCR, multi-image reasoning, and general visual question answering that matches or beats GPT-4V on several public benchmarks, despite running fully offline on a phone.

OpenBMB has kept iterating on the series since, with later versions such as MiniCPM-V 4.5 and 4.6 shrinking the parameter count further while improving accuracy, reflecting a broader trend of closing the gap between small on-device multimodal models and much larger cloud-hosted ones. The models are released under an Apache 2.0 license on Hugging Face and GitHub, aimed at developers who need offline, privacy-preserving multimodal AI on consumer hardware.