中国模型差距美国7个月

Core Insights - A recent report by Epoch AI indicates that Chinese AI models are, on average, 7 months behind their American counterparts, with a minimum gap of 4 months and a maximum of 14 months [1] Group 1: Performance Metrics - The ECI metric developed by Epoch AI measures model performance across various domains such as mathematical reasoning, code writing, and language understanding, integrating results from numerous global AI benchmark tests [3] - From 2024 onwards, the pace of improvement for Chinese large models is expected to accelerate significantly, reducing the gap from 12-14 months in 2023 to approximately 6-8 months, driven by the releases of DeepSeek-V2 and DeepSeek-R1 [3] Group 2: Global Computing Power Landscape - The disparity in the global computing power landscape is notable, with the U.S. controlling about 75% of the world's top GPU cluster performance, while China holds a 15% share as of May 2025 [3] Group 3: Competitive Landscape - The competition between Chinese and American large models is characterized by a divide between open-source and closed-source models, with leading U.S. models like GPT-5, Gemini 3, and Claude 4 being closed-source, while China's DeepSeek and Qwen series adopt varying degrees of open-source strategies [6][7] - The current competitive landscape shows that while U.S. closed-source models continue to set high standards, Chinese firms are leveraging "open-source for ecosystem" strategies to accelerate iteration and enhance competitiveness among global developers and enterprise users [7] Group 4: Future Directions - Both Chinese and American models are approaching performance ceilings after significant growth in parameter scale, introduction of inference modes, and optimization of algorithm architectures, with recent iterations failing to deliver groundbreaking advancements except for Gemini 3 [7] - There is a prevailing sentiment that the era of Scaling Law may be coming to an end, suggesting a shift back to a "research" era where the next disruptive paradigm will define the future of large models [7]