On August 10, CITIC Construction released its latest research report, indicating that the scaling development of cutting-edge AI models has entered a new phase characterized by multi-path parallelism. According to industry estimates, Anthropic’s Mythos 5 and Fable 5 models boast parameter scales of 8 trillion and 5 trillion, respectively, while the Kimi K3 model encompasses a total of 2.8 trillion parameters. Meanwhile, ByteDance is reportedly pre-training a model with an astonishing parameter scale of 10 trillion. During the post-training phase of these models, the focus extends further to managing Agent trajectories involving millions of Tokens, thousands of tool invocations, and complex tasks that span hours. The RSI overview released on July 8 analyzed 1,250 papers, with 74% published in 2026, signaling a significant acceleration in the automation of AI research and development.

CITIC Construction believes that the current competition among models is no longer limited to the expansion of a single parameter. Instead, it has shifted towards the coordinated development of pre-training, reinforcement learning, inference-time computation, RSI (Research and Scientific Innovation), and long-term Agent capabilities. As this trend continues to evolve, the demand for computing power will expand beyond the training phase to encompass inference and Agent execution phases as well. Therefore, CITIC Construction maintains an optimistic outlook on the development prospects of Big Tech’s capital expenditure ecosystem, domestic chip and super node technologies, computing power services, Pre-AI, B-end AI applications, and local inference.