Investment Rating - The report maintains an "Accumulate" rating for the computer industry [6][7]. Core Insights - Perplexity has launched a new version of its Sonar search model, which is optimized for search applications and operates on Cerebras' inference infrastructure. The new model achieves a token decoding speed of 1200 tokens per second, which is over 8.5 times faster than Gemini 2.0 Flash. User evaluations indicate that the performance of the new Sonar model significantly surpasses that of GPT-4omini and Claude 3.5 Haiku, while being comparable to GPT-4o [23][24]. - Huawei's DeepSeek V3/R1 flagship model, with 671 billion parameters, has been fully optimized for deployment on Huawei Cloud's Ascend cloud service, meeting commercial deployment needs [25]. - Tencent's Yuanbao has integrated the DeepSeekR1 model, enabling deep thinking and online search capabilities, enhancing the model's reasoning and response accuracy [26]. - Baidu announced the open-sourcing of its next-generation Wenxin model series 4.5 starting June 30, with a deep search feature available for free from April 1 [27]. Investment Recommendations - The widespread adoption of DeepSeek and improvements in model performance are expected to accelerate the penetration of AI applications. The report is optimistic about the overall industry chain's prosperity and recommends focusing on investment opportunities in the IaaS layer, specifically highlighting companies such as Microsoft (MSFT), Google (GOOGL), Meta (META), and Amazon (AMAZON) [28].
全球AI产业跟踪：DeepSeek：V3/R1满血版上线华为云
Yong Xing Zheng Quan·2025-02-22 12:58