大模型行业专题报告：一文读懂DeepSeek

Investment Rating - The industry rating is "Positive" [6] Core Insights - DeepSeek is a Chinese large model that emphasizes technological innovation and aims to drive the development of the entire ecosystem [1] - DeepSeek has emerged as a disruptive force in the global model market, offering significant impacts in performance, cost, and open-source initiatives [2] - The adoption of DeepSeek's models by major tech companies like NVIDIA, Microsoft, and AWS is accelerating the application of AI technologies [26][30] Summary by Sections 1. Performance and Comparison - DeepSeek R1's performance is comparable to OpenAI's o1 model, achieving a score of 79.8% in the AIME 2024 math benchmark, slightly outperforming OpenAI's 79.2% [2] - The training cost of DeepSeek V3 is approximately $557.6 million, significantly lower than the hundreds of millions required for models like GPT-4 [2][14] 2. Technological Features - DeepSeek employs model distillation techniques to enhance the reasoning capabilities of smaller models, demonstrating superior performance compared to traditional reinforcement learning methods [17][20] - The Janus-Pro framework introduced by DeepSeek enhances multi-modal understanding and generation, improving adaptability across various tasks [24] 3. Market Applications - The B-end applications are expected to benefit the most from the open-source trend and cost reductions, with sectors like customer service and marketing likely to see rapid deployment of AI agents [4][29] - DeepSeek's integration into platforms like Tencent Cloud and Huawei Cloud is expected to facilitate faster development and deployment of AI applications [30] 4. Related Companies - Key companies in AI applications include Kingsoft Office, iFlytek, and Focus Technology, while AI edge companies include Zhongke Chuangda and Hongsoft Technology [5][33]