Abstract: Large model has undergone more than 20 years of invention and optimization. In 2000, Yoshua Bengio pioneered the use of word embeddings and the autoregressive approach for training language models, and proposed self-attention mechanism in 2014. In late 2017, Google invented the Transformer neural network architecture, which relies solely on the attention mechanism. In 2018, OpenAI trained the large language model GPT, which realized intelligence emergence and ignited the intelligence revolution. In 2021, the Beijing Academy of Artificial Intelligence (BAAI) first put forward the concept of “large model” and in 2025, developed the multimodal large model EMU, demonstrating that the autoregressive paradigm practiced by GPT is also applicable to vision and other modalities, and is expected to become a unified path toward generative artificial intelligence.