MiniCPM-V is a family of efficient multimodal large language models from OpenBMB and ModelBest, designed to deliver GPT-4V-class vision and omni-modal capability in models small enough to run on end devices. The latest generation pairs a 35B-A3B mixture-of-experts model and a 1B variant that can run on an iPhone in about 3GB of memory at over 13 tokens per second, alongside MiniCPM-o, a 9B full-duplex omni-modal model that sees, listens and speaks in real time. The project has earned roughly 27,000 stars on GitHub.
MiniCPM-V models accept image, video and text input and are built for strong performance per parameter: high-resolution image understanding, fine-grained OCR, document and chart reading, and video understanding, with techniques that keep visual token counts low so inference stays fast. MiniCPM-o extends the family to streaming, full-duplex interaction across vision, speech and text, enabling real-time multimodal conversation on a single device. The repository provides model weights, inference and fine-tuning code, and deployment paths through llama.cpp, Ollama, vLLM and SGLang, plus demos for phones and edge hardware, making the family a common reference for on-device multimodal AI.
MiniCPM-V is free and open source under the Apache-2.0 license. The models, code and demos can be downloaded and used at no cost, subject to the license terms.
MiniCPM-V is for developers and researchers building multimodal applications that must run on phones, edge devices or modest GPUs, and for anyone who wants capable vision-language and omni-modal models without depending on a cloud API.