互联网：DeepSeek成本和性能双突破，有望加速国内AI进程

Investment Rating - Industry investment rating: Overweight [2] Core Insights - DeepSeek has achieved breakthroughs in both AI model performance and cost, which is expected to accelerate the domestic AI progress [3][4] - The latest AI model DeepSeek-R1 has shown excellent performance in various benchmark tests, particularly in mathematical reasoning and programming tasks, scoring 79.8% in the AIME 2024 math benchmark, slightly higher than OpenAI's o1 model at 79.2%, and 97.3% in the MATH-500 benchmark, outperforming o1's 96.4% [3][4] - DeepSeek has also released the multimodal open-source model Janus-Pro, which outperformed Stable Diffusion and OpenAI's DALL-E 3 in GenEval and DPG-Bench benchmarks [3] - The report highlights the innovative training process and new algorithm applications that DeepSeek has utilized to achieve these advancements [4] Summary by Sections AI Model Performance - DeepSeek-R1's performance is comparable to leading models, indicating significant advancements in AI capabilities [3][4] - The introduction of DeepSeek-V3 showcases a combination of efficient architecture and innovative technologies, including MoE architecture and FP8 mixed-precision training [4] Investment Recommendations - The report suggests that the ongoing decline in unit costs for generative AI will accelerate technology expansion and increase application scenarios [5] - Key companies to watch include Alibaba-W, Tencent Holdings, Baidu Group-SW, and Kingsoft Cloud, as they are expected to benefit from increased demand for cloud infrastructure due to AI adoption [5] - The report also recommends monitoring companies involved in edge AI and SaaS applications, such as Xiaomi Group-W and Kingsoft Office, as they are crucial for AI implementation [5]