集成电路：DeepSeek下载量火爆，端侧爆发有望再加速

Investment Rating - The industry investment rating is "Leading the Market" (维持) [4][11] Core Insights - DeepSeek has rapidly gained popularity with its new open-source model DeepSeek-R1, achieving over 4,000 stars on GitHub shortly after its release [1][2] - The model has demonstrated performance comparable to OpenAI's o1 official version in various tasks, including mathematics, coding, and natural language reasoning [2] - DeepSeek's Janus-Pro multimodal model has outperformed competitors like OpenAI's DALL-E 3 in benchmark tests, achieving accuracy rates of 80% and 84.2% in image generation tasks [3] Summary by Sections Section: DeepSeek's Development - DeepSeek, founded in July 2023, has released multiple open-source models, including DeepSeek Coder, DeepSeek LLM, and DeepSeek-R1, which utilizes reinforcement learning to enhance inference capabilities [2] - The API service for DeepSeek-R1 is priced at 1 RMB per million input tokens (cache hit) and 16 RMB per million output tokens [2] Section: Market Performance - Since the launch of DeepSeek-R1 on January 20, 2025, the app has topped the global download charts on both Apple App Store and Google Play Store, with over 16 million downloads within 18 days across 140 markets [2] Section: Competitive Landscape - Janus-Pro, a new multimodal model from DeepSeek, has shown superior performance in generating images from text compared to existing models, indicating a strong competitive position in the AI market [3] - The integration of knowledge distillation and model miniaturization techniques is maturing, suggesting a comprehensive development of AI across cloud, edge, and endpoint applications [8]