CITIC Securities: Domestic models are rapidly progressing through iteration; continue to pay attention to the domestic large model and computing power industry chain.

CITIC Construction Investment stated that DeepSeek has launched the new generation flagship model DeepSeek-V4-Pro, further enhancing its capabilities in complex reasoning, coding, and agent functionality. Compared to the previously highlighted low-cost, high-efficiency V4-Flash, the V4-Pro focuses more on high-difficulty tasks and agent scenarios, scoring 53 points on the Artificial Analysis Intelligence Index, higher than V4-Flashs score of 40. Zhiyun released its new generation open-source flagship model GLM-5.3 on August 14, continuing to emphasize enhancements in coding, long-range task execution, and agent capabilities. According to Zhiyun, GLM-5.3 has shown an approximately 50% performance improvement on the internal Z.ai Code Bench compared to GLM-5.2, supporting 1 million tokens in context and a maximum output of 128K tokens, further improving in complex software engineering and long-duration autonomous execution tasks. There is a continued recommendation to pay attention to the development of domestic large models and the computing power industry chain. Recently, domestic models such as DeepSeek and Zhiyun have been continuously iterating, accelerating their catch-up with overseas leading levels in reasoning, coding, and agent capabilities, with the model ecosystem and commercialization process speeding up, paving the way for a positive cycle of collaborative development between domestic models and domestic computing power.