China’s AI Rise: In-Depth Analysis of Top Companies and Large Models in Q1 2025 – Based on Artificial Analysis Authority Report

01 Introduction: The New Landscape of AI Competition

In the first quarter of 2025, the global artificial intelligence landscape reached a critical turning point. According to the “State of AI: China Q1 2025” report released by the independent evaluation agency Artificial Analysis, Chinese AI labs have significantly narrowed the gap with the United States in terms of language model intelligence, even surpassing them in some areas. This article delves into the current state and future of China’s AI industry based on core data from the report, analyzing it from three dimensions: technological progress, corporate ecosystem, and hardware limitations.

02 Technological Breakthroughs: The Rise of China’s Large Models

1. Comparison of AI Intelligence Index Between China and the U.S.

The Artificial Analysis Intelligence Index (AAII) is a core metric for measuring the comprehensive capabilities of language models, covering authoritative evaluation sets such as MMLU, GPQA Diamond, MATH-500, and HumanEval. Currently, only the U.S., China, France, and Israel are at the forefront, with China and the U.S. dominating the field. The U.S. no longer holds an absolute advantage, and the two nations are expected to advance side by side in the future.

Leading U.S. Models:

• OpenAI’s GPT-4o+ tops the list with 94 points (released in Q4 2024).
• Anthropic’s Claude 3.5 Sonnet+ (89 points) and Google’s Gemini 2.0 Flash+ (84 points) follow closely.
• XAI’s Grok3 shows strong momentum, entering the top tier of large models, with reasoning performance comparable to OpenAI’s o3.

Leading Chinese Models:

• DeepSeek’s R1 model ranks third globally with 89 points (released in Q4 2024), only 5 points behind GPT-4o.
• Alibaba’s Qwen 2.5 Max+ (82 points), Baichuan Intelligence’s M1-Preview+ (83 points), Zhipu AI’s GLM-Zero-Preview (81 points), Minimax’s Text-o1, and Moonshot’s Kimi1.5 all enter the top tier.

2. Reasoning Models: From Followers to Equals

In Q3 2024, OpenAI introduced the first reasoning model with “thinking” capabilities (o1 series), but Chinese companies responded impressively:

• DeepSeek released the R1 model+ (89 points) within 3 months, matching the performance of GPT-4o.
• By Q1 2025, seven Chinese labs had launched reasoning models, including Moonshot (Kimi k1.5, 87 points) and Zhipu AI (GLM-Zero-Preview, 81 points).

3. Leading Advantages in Open-Source Models

Chinese companies have made significant strides in open-source initiatives:

• Open-source models like DeepSeek V3+ (79 points) and Alibaba’s Qwen 2 Instruct 72B (80 points) perform close to the cutting-edge closed-source models.
• The report notes that the developer ecosystem for Chinese open-source models has reached 700,000 (e.g., Zhipu AI’s ChatGLM), far surpassing similar U.S. companies.

03 Corporate Ecosystem: Dual Drivers of Giants and Startups

1. Tech Giants: Full-Modal Layouts

China’s leading tech companies have established complete ecosystems of “models + applications + cloud services”:

Other domestic giants like Meituan, Xiaomi, Kunlun Wanwei, iFlytek, and Qihoo 360 have also joined the competition.

2. AI Unicorns: Breakthroughs in Vertical Fields

Startups have shown remarkable innovation in niche sectors:

• DeepSeek: Originating from a quantitative trading team, its R1 model (89 points) excels in finance and code generation.
• Moonshot: Launched the Kimi model (87 points) with a 2-million-word context window, focusing on long-text reasoning. It is predicted to become a top-tier model in the future.
• Zhipu AI: GLM-Zero-Preview (81 points) has been adopted by over 300 enterprise clients in healthcare and legal fields.
• Baichuan Intelligence: M1-Preview (83 points) focuses on medical AI, collaborating with Peking Union Medical College Hospital to develop diagnostic assistance systems.

Other notable startups include MiniMax, 01.AI, and StepFun.

Most unicorn AI companies are backed by major investors like Alibaba, Tencent, Xiaomi, Microsoft, Gaorong, and Xiaohongshu. DeepSeek, however, has no external giant funding, though its valuation is estimated at around USD 10 billion.

3. Government and Capital Support

• Policy Support: China’s “14th Five-Year Plan” AI special project has invested over USD 15 billion, focusing on chips, frameworks, and open-source communities.
• Capital Enthusiasm: In 2024, China accounted for 38% of global AI venture capital, second only to the U.S. (45%).

04 Hardware Limitations: Challenges and Opportunities in Domestic Substitution

1. Escalating U.S. Export Controls

Since 2022, the U.S. has progressively tightened restrictions on AI chip exports to China:

• October 2022: A100/H100 banned; NVIDIA released “neutered” versions A800/H800.
• October 2023: A800/H800 also banned; NVIDIA forced to release H20 with only 148 TFLOPs (15% of H100).
• January 2025: AI diffusion rules took effect, restricting third-country resales. Huawei’s Ascend 910B+ became the preferred domestic alternative.

2. Progress in Domestic Chips

• Huawei Ascend 910B: 256 TFLOPS (FP16), close to A100 (312 TFLOPS), deployed in over 20 intelligent computing centers.
• Cambricon MLU370+: Supports 128-way interconnection, scaled on Alibaba Cloud.
• Moore Thread MTT S4000+: First domestic GPU supporting FP8 precision, compatible with CUDA ecosystem.
• Model Lightweighting: Alibaba Qwen 2.5 released a 7B version (Intelligence Index 72), reducing inference energy consumption by 60%.
• Heterogeneous Computing: Tencent’s Hunyuan model uses a “CPU+NPU+FPGA” hybrid architecture, cutting training costs by 35%.

05 Future Outlook: Opportunities and Concerns

1. Technological Trends

• Multimodal Fusion: ByteDance’s Doubao 1.5 Pro supports text, image, and 3D generation (AAII 80 points).
• Autonomous Frameworks: Huawei’s MindSpore and Baidu’s PaddlePaddle have surpassed 2 million global developers.
• Agent Workflows: ByteDance’s Coze platform will empower industries like writing, graphics, video, social media, e-commerce, and SaaS.
• Vertical-Specific Models: Specialized large models will flourish, potentially creating hidden giants in niche fields.

2. Challenges and Risks

• Computing Bottlenecks: Chinese AI chips still lag behind global leaders by 1-2 generations.
• Data Compliance: International data flow restrictions are tightening, pushing some companies to use synthetic data (e.g., DeepSeek’s SynthQA dataset).

3. Competitive Landscape Predictions

• End of 2025: China may surpass the U.S. in open-source models and vertical applications.
• 2026: If 3nm domestic GPUs are mass-produced, the computing gap could narrow to within six months.

06 China’s AI “Golden Window”

From technological catch-up to ecosystem innovation, China’s AI industry is reshaping the global competitive landscape at an astonishing pace. Despite hardware limitations being the biggest constraint, flexible corporate strategies and policy support have opened new possibilities. The key to the future lies in transforming engineering advantages into genuine technological originality.