MiniCPM-o 4.5 Documentation

MiniCPM-o 4.5 PyTorch Simple Demo is an official inference and demonstration system provided by the model training team. It uses a PyTorch + CUDA inference backend to showcase the full audio-video omnimodal full-duplex capabilities of MiniCPM-o 4.5 in a transparent and concise manner.

The system supports four interaction modes (Turn-based Chat, Half-Duplex, Omnimodal Full-Duplex, Audio Full-Duplex), sharing a single model instance with millisecond-level hot-switching.

Quick Start

• Architecture — System architecture, mode topology, and request flow
• Model — Multimodal model internals and data flow
• API Reference — Interface definitions and usage
• Deployment — Requirements, configuration, and deployment steps