openbmb/minicpm-o

MiniCPM-o is a series of compact, open-source multimodal AI models designed to run efficiently on devices like phones and laptops. Multimodal means the models can process multiple types of input simultaneously, images, video, audio, and text, and produce text and speech responses. The flagship model, MiniCPM-o 4.5, has 9 billion parameters and is designed to match the capability of Google's Gemini 2.5 Flash while being small enough to deploy locally. Its headline feature is full-duplex multimodal live streaming, meaning the model can see, listen, and speak all at the same time without each operation blocking the others. You can have a real-time conversation where the model watches your camera feed, hears your voice, and responds with speech, all simultaneously, like a video call with an AI. Features include voice cloning, bilingual real-time speech conversation, optical character recognition in images, and proactive interaction (the model can initiate reminders on its own). A companion model, MiniCPM-V 4.0, focuses on image understanding at just 4 billion parameters and outperforms much larger models on certain benchmarks. You would use MiniCPM-o when building on-device AI assistants, accessibility tools, or real-time interactive applications where sending data to a cloud server is impractical or undesirable. The tech stack is Python, with support for deployment via llama.cpp, Ollama, and vLLM.