DeepSeek R1 更新，Agent 生态有望加速演进

Investment Rating - Investment recommendation: Outperform the market (maintained) [7] Core Insights - From a model perspective, the DeepSeek R1 model, which is set to be open-sourced in early 2025, is expected to narrow the performance gap between domestic AI models and their overseas counterparts. The next-generation R2 model's release schedule is worth monitoring [4][9] - 2025 is anticipated to be the year of AI application implementation, with AI Agents recognized as a significant direction for AI applications globally. Domestic companies are expected to leverage enhanced model capabilities and a vast C-end user ecosystem to foster widespread AI applications [4][9] Summary by Sections Model Performance - The latest version of DeepSeek R1 has shown significant improvements in deep thinking capabilities, achieving a mathematical accuracy rate of 87.5% in the AIME2025 test, up from 70% in the previous version. The average token usage per question increased from 12K to 23K, indicating deeper reasoning [8][9] - The model's text processing capabilities have also improved, with a 45-50% reduction in "hallucination" rates, leading to more accurate and reliable outputs in various scenarios [8][9] Application Landscape - The report highlights that AI Agents are becoming a consensus among global tech giants, with significant developments in both overseas and domestic markets. Companies like Tencent, Alibaba, and ByteDance are expected to launch Agent products, capitalizing on improved model capabilities and user ecosystems [4][9][10] - The report suggests monitoring companies that have early layouts in AI models and applications, including Kaiying Network, Kunlun Wanwei, Giant Network, and others [4][10]