According to CoinMeta, a research report from CITIC Construction Investment indicates that the cutting-edge model scaling has entered a multi-path parallel development phase. The models Anthropic Mythos5 and Fable5 are estimated to have 8 trillion and 5 trillion parameters respectively, while Kimi K3 has a total of 2.8 trillion parameters. ByteDance is reportedly preparing to pre-train models with up to 10 trillion parameters. The RSI overview, released on July 8th, covers 1,250 papers, 74% of which were published in 2026. The development of automation for AI has significantly accelerated. CITIC Construction Investment believes that model competition is shifting from a focus on individual parameters to a coordinated evolution of capabilities such as pre-training, reinforcement learning, and inference computing. The demand for computing power will expand from training to inference and agent execution. They continue to be optimistic about the ecosystem of large companies Capex, domestic chips and super nodes, computing power services, Pre-AI, B-side AI applications, and local inference.