人工智能行业点评报告：DeepSeek领衔，国产大模型竞争力加速迭代

Investment Rating - The industry investment rating is "Positive" (maintained) [7] Core Insights - DeepSeek has released the DeepSeek-R1 model and multi-modal frameworks Janus-Pro and JanusFlow, showcasing excellent performance with fewer parameters and lower costs, indicating a competitive edge in the domestic AI model landscape [1][2] - The DeepSeek-R1 model, with minimal labeled data, performs comparably to OpenAI's o1 model in various tasks, and its API service pricing is significantly lower, reducing costs by over 90% compared to competitors [2] - The Janus-Pro framework enhances model adaptability and performance across tasks, utilizing innovative encoding methods to understand and generate multi-modal content [3] - Alibaba has launched the Qwen2.5-VL model, which can automate operations on computers and mobile devices, improving image, text, and video understanding capabilities [4] - Doubao has introduced a real-time voice model that excels in emotional understanding and expression, achieving near-human speech quality in Chinese contexts [5] Summary by Sections - DeepSeek Model Launch: The release of DeepSeek-R1 and its integration into major cloud platforms like Tencent Cloud and Huawei Cloud is expected to accelerate commercial applications [2] - Janus-Pro Framework: Janus-Pro's architecture allows for enhanced understanding of images and scene relationships, achieving impressive results with a relatively small training setup [3] - Alibaba's Qwen2.5-VL: This model's ability to function as a visual agent marks a significant advancement in AI capabilities, particularly in video comprehension [4] - Doubao's Voice Model: The model's features include low latency and the ability to interrupt conversations, showcasing advancements in voice interaction technology [5]