Survey of Major Chinese AI Large Language Models: Technologies, Innovations, and Comparative Evaluation

Abstract With the rapid development of artificial intelligence, deep learning has become the core driving force behind natural language processing, computer vision, and multimodal applications. This report systematically reviews the key technologies, innovations, and performance of the main Chinese AI large models such as GPT‑style series, Qwen, DeepSeek, and Kimi, and provides a comprehensive comparison.

1. Introduction Deep learning, from perceptrons to the Transformer era, has enabled breakthroughs in image recognition, NLP, medical diagnosis, and autonomous driving. Recent Transformer‑based large language models (LLMs) have achieved remarkable progress in language generation, reasoning, multilingual support, and multimodal fusion.

2. Evolution of Deep Learning The section recounts the history from the perceptron (1958) to multilayer networks with back‑propagation, convolutional and recurrent networks, the attention mechanism (2014) and the Transformer (2017), and finally the emergence of GPT‑style generative pre‑training models.

3. Chinese Innovations: DeepSeek and Kimi DeepSeek introduces a Mixture‑of‑Experts (MoE) architecture, Multi‑Head Latent Attention, Multi‑Token Prediction, an auxiliary‑loss‑free strategy, and FP8 mixed‑precision training, achieving higher efficiency and performance. Kimi focuses on long‑context scaling (up to 128 k tokens), improved policy optimization based on online mirror descent, a simplified reinforcement‑learning framework, and multimodal training, delivering state‑of‑the‑art results on benchmarks such as AIME 2024 and MATH‑500.

4. Qwen2.5 Technical Highlights Qwen2.5 expands pre‑training data to 18 trillion tokens, employs a refined data‑filtering pipeline, and adopts multi‑level reinforcement learning (offline DPO and online GRPO). It supports both dense and MoE configurations, introduces YARN and dual‑cache attention for up to 1 million token contexts, and integrates multimodal fusion capabilities.

5. Comparative Evaluation DeepSeek‑V3, DeepSeek‑R1, Qwen2.5, and Kimi 1.5 are compared in architecture, training cost, inference speed, multilingual ability, multimodal performance, and benchmark scores (MMLU, MATH‑500, AIME, etc.). DeepSeek excels in mathematical and code tasks, Qwen2.5 offers strong multilingual and cost‑effective performance, while Kimi shines in multimodal and long‑context scenarios.

6. International Trend Comparison Foreign models continue to grow in scale (GPT‑3, LLaMA, Gemini) and explore multimodal fusion and reinforcement learning. Chinese models differentiate themselves through efficient training (MoE, FP8), deeper reinforcement‑learning pipelines, extensive multimodal integration, and breakthroughs in ultra‑long context handling.

7. Conclusion and Outlook Chinese LLMs demonstrate notable innovations in efficiency, multimodality, and long‑context processing. Future directions include further training‑efficiency improvements, deeper multimodal fusion, enhanced long‑text capabilities, and better model interpretability.