CMSC: Knowledge Atlas (02513) Reasoning Efficiency Breakthrough Further Enhances Commercial Prospects Rating "Hold"

CMSC International released a research report stating that the reasoning efficiency of KNOWLEDGE ATLAS (02513) has further improved, enhancing its commercial prospects. It has been given a "hold" rating and the target price is currently under review. Last Friday (22nd), the high-speed version API GLM-5.1 was released with an output speed of 400 tokens/s, significantly increasing model output speed and surpassing the global maximum speed limit for API from major model manufacturers. This is the first time that flagship capabilities and extreme low latency have been simultaneously brought into a production environment in a domestic large model. The report mentions that the new architecture reduces inference costs. Last Thursday (21st), the next-generation inference network architecture ZCube was released, with the landing of production clusters increasing throughput by over 15%, reducing TTFT P99 by 40.6%, and reducing network hardware costs by 33%. CMSC stated that in an Agent scenario with tens of rounds of calls, speed differences are exponentially amplified. The high-speed version enables previously unattainable product forms such as real-time collaboration, 3D interactive modeling, and real-time tool generation due to latency. Reasoning efficiency is the core leverage of the MaaS business model: a combination of 400 TPS + ZCube (15% increase in throughput, 33% reduction in network costs), while optimizing both ends of the flywheel - increasing the rate of intelligent supply and reducing the unit token production cost. If the high-speed version is fully launched, it is expected to further improve ARR expectations and profit margin expectations. The company continues to have a positive outlook on KNOWLEDGE ATLAS, focusing on "pursuing the upper limit of model intelligence" as its core. The continuous deepening of programming capabilities, technological barriers, and the commercial closed loop will continue to enhance the model's pricing power. Future key areas of focus include (1) inclusion in the Hang Seng Index/Hang Seng Technology Index in early June and subsequent inclusion in the Hong Kong Stock Connect; (2) continuous iteration of the GLM model; (3) growth of open platform ARR and improvement in profit margins; (4) the progress of listing on the Science and Technology Innovation Board in the second half of the year; (5) the improvement and progress of domestic chip supply and compatibility. Major risks include (1) a large shareholder reducing holdings after the lock-up period ends; (2) model iteration falling short of expectations and increased competition in models; (3) exceeding anticipated investment.