DeepSeek升级R1，称性能“接近OpenAI”

Core Insights - DeepSeek has upgraded its AI model "R1," achieving a correct rate of 87.5% in solving problems from the 2025 International Mathematical Olympiad US Qualifying Exam, an improvement from the previous 70% [1] - The latest logic reasoning model "o3" from OpenAI, released in April, has a correct rate of 88.9% in the same test [1] - DeepSeek's R1 model is designed for logical reasoning and excels in solving complex problems through extended thinking time, positioning itself as a competitor to OpenAI and Google in the AI development space [1] Group 1 - DeepSeek's R1 model performance has improved due to increased computational resources and reduced error rates in generated answers [2] - The initial version of R1 was released in January, demonstrating performance comparable to OpenAI's earlier model "o1" at a lower cost, indicating strong AI development capabilities in China [2] - The open-source nature of DeepSeek's AI model allows external developers to utilize and modify it freely, although concerns about data leakage have been raised due to its Chinese origin [2] Group 2 - In response to DeepSeek's advancements, OpenAI has announced plans to open-source its logic reasoning model to compete with R1 [3]