OpenBMB/MiniCPM-V

A Pocket-Sized MLLM for Ultra-Efficient Image and Video Understanding on Your Phone

What it solves

MiniCPM-V and MiniCPM-o provide high-performance multimodal large language models (MLLMs) that are small enough to be deployed efficiently on mobile devices (iOS, Android, and HarmonyOS) without sacrificing significant capability in image, video, and audio understanding.

How it works

  • MiniCPM-V 4.6 uses a 1.3B parameter architecture based on SigLIP2-400M and Qwen3.5-0.8B. It employs an intra-ViT early compression technique to reduce visual encoding computation costs by over 50% and supports mixed 4x/16x visual token compression to balance accuracy and speed.
  • MiniCPM-o 4.5 is a 9B parameter omnimodal model capable of real-time, full-duplex interaction. This means it can process streaming video and audio inputs while simultaneously generating speech and text outputs without blocking, enabling proactive interactions.

Who it’s for

  • Mobile Developers: Those looking to integrate advanced vision-language or omnimodal capabilities directly onto smartphones.
  • AI Researchers: Users interested in efficient MLLM architectures and edge-deployment techniques.
  • Local LLM Users: People using frameworks like Ollama, vLLM, or llama.cpp to run multimodal models on their own hardware.

Highlights

  • Edge-Optimized: Specifically designed for deployment on iOS, Android, and HarmonyOS with open-source adaptation code.
  • High Efficiency: MiniCPM-V 4.6 achieves ~1.5x token throughput compared to Qwen3.5-0.8B despite having stronger capabilities.
  • Omnimodal Interaction: MiniCPM-o 4.5 supports simultaneous seeing, listening, and speaking in real-time.
  • Broad Framework Support: Compatible with SGLang, vLLM, llama.cpp, Ollama, SWIFT, and LLaMA-Factory.

Related

  • Project
    OpenBMB/MiniCPM-V-Apps

    A collection of on-device mobile demos for the MiniCPM-V family of multimodal models, enabling local AI inference on iOS, Android, and HarmonyOS.

  • Project
    OpenBMB/MiniCPM-o-Demo

    A demo system for MiniCPM-o 4.5 that enables real-time, full-duplex omnimodal interaction, allowing the AI to see, listen, and speak simultaneously.

  • Project
    OpenSQZ/MiniCPM-V-CookBook

    A comprehensive implementation guide and deployment cookbook for the MiniCPM series, enabling the integration of vision, speech, and text AI capabilities across edge devices and enterprise servers.

  • Project
    UbiquitousLearning/mllm

    A fast and lightweight multimodal LLM inference engine for mobile and edge devices that optimizes model execution across CPUs, GPUs, and NPUs.

  • Project
    jingyaogong/minimind-v

    A minimal implementation of a Vision Language Model (VLM) that allows users to train a 65M parameter multimodal model on a single consumer GPU in just 2 hours.