传媒：DeepSeek带来的AI变革

Investment Rating - The industry rating is "Outperform the Market" [7][16] Core Insights - DeepSeek has rapidly gained popularity, topping global download charts, with significant user growth following the release of V3 and R1 [2][3] - The architecture innovations in DeepSeek, including MOE and MLA, significantly enhance model performance and reduce memory usage [4] - The competition in AI applications and terminals is intensifying, with DeepSeek driving down costs and pushing for product capability competition [5] - The increase in user scale and frequency of use is expected to drive up inference costs, despite historical trends of decreasing computational costs [6] Summary by Sections User Growth - DeepSeek's web version has seen heavy user engagement since the launch of V3 on December 26, 2024, and further growth after the R1 release on January 20, 2025 [2] - On the app side, DeepSeek matched GPT-o1 in downloads shortly after R1's launch, reaching the top of the download charts in 168 countries by January 30, 2025 [3] Architectural Innovations - The MOE architecture in DeepSeek allows for a significant reduction in the number of active parameters during inference, utilizing 1 shared expert and 256 routing experts per layer [4] - The MLA architecture reduces memory usage by caching fewer low-rank vectors during inference, enhancing efficiency [4] - DeepSeek's training approach using low-precision FP8 has yielded positive results, challenging conventional training methods [4] Market Impact - The advancements in DeepSeek are expected to accelerate the development of large models across the industry, enhancing competitiveness against established models like GPT4o and GPT-o1 [5] - The anticipated rise in inference costs is linked to increased user engagement, with historical data indicating a 1000-fold decrease in computing costs over 60 years, contrasted with a 100-fold increase in total spending [6] Investment Recommendations - The report expresses optimism about the rise of Chinese large models, particularly in AI applications and terminals, predicting significant growth in inference costs within the global AI competitive landscape [6]