计算机行业周报：国内模型开启世界级竞赛

Investment Rating - The report maintains an "Increase" rating for the industry [5] Core Insights - The launch of the open-source DeepSeek-R1 model has created significant competition for OpenAI's models, showcasing comparable performance in various tasks [12][16] - The advancements in domestic AI models, such as DeepSeek and Doubao, are expected to accelerate the practical application of AI in China, leading to lower costs and higher efficiency for enterprises [44][45] - The report emphasizes the importance of algorithmic innovation and engineering optimization in achieving high performance at lower costs, highlighting China's potential to catch up with global leaders in large model technology [20][29] Summary by Sections DeepSeek Model Launch - DeepSeek-R1 was released on January 20, 2025, and is open-sourced under the MIT License, allowing users to distill other models using its training [12] - The model's performance is on par with OpenAI's o1 version, achieving significant improvements in reasoning tasks with minimal labeled data [12][20] Technical Innovations - DeepSeek's innovations include a reinforcement learning-driven approach without supervised fine-tuning, which significantly enhances reasoning capabilities [20][22] - The introduction of the GRPO algorithm reduces training costs, and a rule-based reward system is employed to evaluate model accuracy [22][23] - The architecture combines multi-head attention mechanisms and expert mixture models for efficient training and inference [23][26] Doubao Model Developments - Doubao's real-time voice model was launched on January 20, 2025, demonstrating superior emotional understanding compared to GPT-4O [30][33] - The Doubao-1.5-pro model, released on January 22, 2025, shows significant performance improvements across various benchmarks, utilizing a sparse MoE architecture [35][36] Investment Opportunities - The report identifies several investment opportunities arising from advancements in domestic AI models, including collaborations with internet giants and the development of AI agents for various applications [44][49] - The increasing efficiency of computational resources is expected to lower barriers for new entrants in the large model industry, creating a favorable environment for growth [48] - The report suggests a diverse range of investment targets, including AI software companies, internet giants, and military AI applications [55][56]