Abstract: With the rise of the Industry 5.0 paradigm, the traditional Industrial Internet is accelerating its evolution towards the Industrial Internet of Intelligence, which is human-centric, knowledge-automated and resilient. In response to the long-existing problems in industrial systems, such as the semantic interaction barriers of cross-modal heterogeneous data, the computational capacity bottleneck at the edge, insufficient flexibility in control execution and the hallucinations of generative artificial intelligence, we systematically construct a "five-dimensional integrated" architecture of the Industrial Internet of Intelligence based on the Artificial Societies, computational experiments, parallel execution (ACP) theory, which is empowered by large models. We review the key technological evolutions, empowerment mechanisms and industrial application scenarios of large models across five dimensions of perception, communication, computation, control, and intelligence. For perception, we dissect the technological evolution of native multimodal architectures, long-sequence time-series forecasting, and 3D spatial perception. For communication and computation, we detail mechanisms that include semantic communication, federated learning, cloud-edge collaborative computing, as well as model distillation and quantization. For control and intelligence, we discuss technologies such as cognitive agent networks, industrial knowledge graphs, and generative digital twins. Finally, we analyze the challenges faced by large-model-driven Industrial Internet of Intelligence and propose prospects for future industrial architectures.