CMSC International released a research report stating that KNOWLEDGE ATLAS (02513) has achieved a breakthrough in inference efficiency, further improving its commercial prospects. It maintains an "Add" rating and is currently reviewing the target price. KNOWLEDGE ATLAS launched its GLM-5.1 High-Speed API last Friday (22nd), achieving an output speed of 400 tokens per second. This significant enhancement in model output speed sets a new global benchmark for API speed among major model providers. It marks the first time a domestic large model has simultaneously introduced flagship capabilities and extremely low latency into a production environment. The report notes that the new architecture reduces inference costs. The company released its next-generation ZCube inference network architecture last Thursday (21st), with a thousand-card production cluster now operational. This has increased throughput by over 15%, reduced TTFT P99 by 40.6%, and cut network hardware costs by 33%. CMSC indicated that in Agent scenarios involving dozens of rounds of calls, speed differences are amplified exponentially. The high-speed version enables product forms previously impossible due to latency, such as real-time collaboration, 3D interactive modeling, and real-time tool generation. Inference efficiency is a core lever for the MaaS business model: the combination of 400 TPS and ZCube (15% higher throughput, 33% lower network costs) optimizes both ends of the flywheel—increasing the rate of intelligent supply and reducing the production cost per token. If the high-speed version is fully deployed, it is expected to further boost expectations for Annual Recurring Revenue (ARR) and margin improvement. The firm remains optimistic about KNOWLEDGE ATLAS's core focus on "pursuing the upper bound of model intelligence," continuously deepening its technical moat and commercial closed loop in programming capabilities, thereby enhancing its model pricing power. Key future focuses include: (1) Potential inclusion in the Hang Seng Index/Hang Seng Tech Index in early June and subsequent inclusion in the Stock Connect program; (2) Continuous iteration of the GLM model; (3) Growth in open platform ARR and margin improvement; (4) Progress towards a STAR Market listing in the second half of the year; (5) Improvements in the supply and adaptation progress of domestic chips. Major risks include: (1) Shareholder减持 after the lock-up period ends; (2) Model iteration falling short of expectations and intensifying model competition; (3) Expenditure exceeding expectations.