NetGPT is an AI-native network architecture designed to provide personalized generative services beyond traditional AI applications. The paper introduces NetGPT, which synergizes large language models (LLMs) at the edge and cloud based on their computational capabilities. Edge LLMs can leverage location-based information for personalized prompt completion, enhancing interaction with cloud LLMs. The feasibility of NetGPT is demonstrated through low-rank adaptation-based fine-tuning of open-source LLMs, such as GPT-2-base and LLaMA, and comprehensive comparisons with alternative cloud-edge collaboration and cloud-only techniques. The paper highlights essential changes required for an AI-native network architecture, emphasizing deeper integration of communications and computing resources and careful calibration of logical AI workflows. NetGPT offers benefits such as unified solutions for intelligent network management and orchestration, as edge LLMs can predict trends and infer intents. The paper discusses the implementation of NetGPT, including the DNN structures of LLMs at the edge and cloud, low-rank adaptation and cloud LLM fine-tuning, edge LLM fine-tuning, and performance showcases. It also explores the extension of NetGPT to large multi-modal models (LMMs) and the AI-native network architecture towards NetGPT. The architecture requires converged C&C resource management, data processing and privacy protection, personalized profiling, and logical AI workflows. NetGPT provides a unified solution for network management and orchestration, including popularity prediction and intent inference. The paper concludes with future research directions, including the challenges of deploying NetGPT, such as implementing inference and fine-tuning at terminals, adapting to dynamic wireless environments, improving LLM rigor, and evaluating model outputs. NetGPT represents a promising AI-native network architecture for provisioning beyond personalized generative services.NetGPT is an AI-native network architecture designed to provide personalized generative services beyond traditional AI applications. The paper introduces NetGPT, which synergizes large language models (LLMs) at the edge and cloud based on their computational capabilities. Edge LLMs can leverage location-based information for personalized prompt completion, enhancing interaction with cloud LLMs. The feasibility of NetGPT is demonstrated through low-rank adaptation-based fine-tuning of open-source LLMs, such as GPT-2-base and LLaMA, and comprehensive comparisons with alternative cloud-edge collaboration and cloud-only techniques. The paper highlights essential changes required for an AI-native network architecture, emphasizing deeper integration of communications and computing resources and careful calibration of logical AI workflows. NetGPT offers benefits such as unified solutions for intelligent network management and orchestration, as edge LLMs can predict trends and infer intents. The paper discusses the implementation of NetGPT, including the DNN structures of LLMs at the edge and cloud, low-rank adaptation and cloud LLM fine-tuning, edge LLM fine-tuning, and performance showcases. It also explores the extension of NetGPT to large multi-modal models (LMMs) and the AI-native network architecture towards NetGPT. The architecture requires converged C&C resource management, data processing and privacy protection, personalized profiling, and logical AI workflows. NetGPT provides a unified solution for network management and orchestration, including popularity prediction and intent inference. The paper concludes with future research directions, including the challenges of deploying NetGPT, such as implementing inference and fine-tuning at terminals, adapting to dynamic wireless environments, improving LLM rigor, and evaluating model outputs. NetGPT represents a promising AI-native network architecture for provisioning beyond personalized generative services.