Abstract: With the rapid development of artificial intelligence technology, large models have become the core driving force behind breakthroughs in natural language processing, multimodal understanding, and complex reasoning tasks. This paper systematically reviews the development and technological evolution of large models, categorizing them into foundational language models, multimodal large models, and reasoning large models. It provides an in-depth analysis of model architecture designs centered around Transformer and key technological improvements. In the field of language models, we focus on innovations in pretraining paradigms, tracing the progress from BERT and the GPT series to models such as LLaMA and GLM-130B, as well as the advancements in task generalization enabled by instruction-tuned models like T0 and Flan-LM. For multimodal models, we examine technical breakthroughs in cross-modal alignment and joint representation learning, exemplified by models such as GPT-4 and Qwen-VL. Regarding reasoning large models, we explore cases such as OpenAI-o1 and DeepSeek-R1, highlighting the synergy between mixture-of-experts (MoE) systems and reinforcement learning optimization strategies. This paper presents a comprehensive overview of the technical framework of DeepSeek-R1, covering its DeepSeekMoE architecture, multi-head latent attention (MLA) mechanism, and group relative policy optimization methods. Additionally, we validate the effectiveness of its three-stage training process in extending long-text capabilities and refining reasoning alignment. Research findings indicate that the continuous scaling of model size must be coupled with training efficiency optimization, knowledge integration strategies, and human-AI alignment techniques. Furthermore, multi-stage reinforcement learning and rejection sampling mechanisms can significantly enhance reasoning reliability in complex scenarios.