计算机行业重大事项点评：DeepSeek: V3.2重塑开源模型性能边界

Investment Rating - The report maintains a "Recommendation" rating for the industry, expecting the industry index to rise more than 5% over the next 3-6 months compared to the benchmark index [39]. Core Insights - DeepSeek-V3.2 has achieved leading reasoning performance among open-source models, reaching a score of 93.1% in the AIME 2025 mathematics competition, closely following GPT-5 at 94.6% [10][14]. - The DeepSeek-V3.2-Speciale version focuses on extreme reasoning and has won gold medals in prestigious competitions such as IMO 2025 and ICPC World Finals 2025, showcasing its advanced capabilities [13][14]. - The introduction of DSA technology in DeepSeek-V3.2 significantly enhances both efficiency and performance, reducing computational complexity from O(L²) to O(Lk) [19][21]. - DeepSeek-V3.2 is the first model to integrate deep thinking with tool usage, improving its reasoning generalization ability through innovative training methods [21][25]. Summary by Sections DeepSeek-V3.2: Redefining Open-Source Model Performance - DeepSeek-V3.2 was officially released on December 1, 2025, and aims to balance reasoning ability with output length, making it suitable for everyday tasks [10]. - The model's performance in benchmark tests is comparable to that of leading proprietary models, demonstrating its competitive edge in the open-source domain [10][12]. Investment Recommendations - The report suggests focusing on specific sectors within AI, including domestic computing power and enterprise-level services, highlighting companies such as Cambricon, Alibaba, and Hikvision [27][28]. - Various application scenarios are identified, including industrial, military, medical, and legal sectors, with specific companies recommended for investment [28].