Large Models: Concepts, Principles, Classifications and Applications

In the current wave of digital transformation, large models have emerged as a breakthrough technology that reshapes how we live and work, driving economic growth, social governance, and innovation. They are massive AI models built on deep learning, featuring billions to trillions of parameters, extensive training data, and high computational resource demands.

Large models, exemplified by OpenAI's GPT‑3 (175 billion parameters) and GPT‑4 (over 1.8 trillion parameters), as well as Alibaba's M6 (10 trillion parameters), require distributed training and specialized hardware acceleration. Their core capability stems from the Transformer architecture, which processes text as tokens, maps each token to a vector, and uses self‑attention to capture relationships across the entire sequence.

DeepSeek is a key player in large‑model development. Its DeepSeek‑V3, released on 2024‑12‑26, matches or exceeds the performance of top closed‑source models such as GPT‑4o while costing only a fraction of the training budget. Subsequent releases like DeepSeek‑R1 and the multimodal Janus‑Pro further demonstrate advances in mathematical reasoning, code generation, and text‑to‑image synthesis.

Large models can be categorized by the type of data they handle: (1) Language models (LLMs) for natural‑language processing, (2) Vision models for image analysis, and (3) Multimodal models that integrate text, images, audio, etc. Representative products include GPT series, Bard, Wenxin Yiyan, VIT, Wenxin UFO, and DALL‑E.

Common AIGC tools built on these models—such as OpenAI's ChatGPT, DeepSeek, Baidu's Wenxin Yiyan, iFlytek's Xinghuo, Alibaba's Tongyi Qianwen, Huawei's Pangu, ByteDance's Doubao, and Kimi—offer capabilities like text generation, knowledge Q&A, logical reasoning, and content creation, serving applications from writing assistance to intelligent customer service.

The report also provides a partial PPT (141 pages) illustrating these concepts, with instructions to obtain the full editable PPT by following the "DBLAB Database Laboratory" WeChat account and replying with "Xiamen University".