Fangzheng Securities: Domestic flagship models are being released intensively, and the industrial chain boom has reached another level

The Zhitong Finance App learned that Fangzheng Securities released a research report saying that recently domestic flagship models have been intensively released, domestic big models and overseas SOTA have entered the same competitive range, and the boom in the AI industry chain has reached another level. Smart Spectrum, Dark Side of the Moon, and Ali have been on the same stage for several days, and GLM5.2, Kimi K3, and Qwen3.8-Max have competed on the same stage. Looking at the computing power industry chain, model parameters and inference requirements have jumped at the same time, and demand for training and reasoning computing power has been rigidly expanding over a long period of time. The domestic model opened the ceiling of global demand, and commercialization accelerated. In terms of network security, large-scale agent deployment and data flow are driving security protection requirements.

Fangzheng Securities's main views are as follows:

Kimi K3: The world's largest parametric open source weight model

On July 17, Kimi K3 was released on Dark Side of the Moon. It uses a 2.8 trillion parameter MoE architecture (896 experts, 16 activations per token), which is currently the open source model with the largest parameters in the world; it has 1M context and native visual understanding, and uses Kimi Delta Attention hybrid linear attention in terms of architectural innovation. In terms of performance, K3 is in the comparable range of Claude Fable 5 and GPT-5.6 Sol in several mainstream programming and agent benchmarks. Only two days after the release of this version, new user subscriptions were suspended due to a surge in user requests and the computing power cluster approaching the carrying limit. Demand elasticity on the inference side greatly exceeded expectations, which is the most direct sign of the computing power boom.

Smart Spectrum GLM-5.2: open source benchmark, domestic base

Zhiguo GLM-5.2 was fully open sourced using the MIT protocol in mid-June, and the Artificial Analysis Open Source Intelligence Index reached 51 points, once leading the global open source model; GLM-5.2 was specially designed for long-range task capabilities, has stronger sensory and more practical coding capabilities, and is fully compatible with domestic chips such as Huawei. In terms of pricing, the GLMCoding Plan Lite monthly subscription package was lowered to 49 yuan, and the API output unit price was about 28 yuan/million tokens, greatly lowering the usage threshold for developers. Combining smart spectra to land in Hong Kong stocks, the company formed an “open source ecology+localization base+capitalization” three-wheel drive.

Ali Qwen3.8-Max: 2.4 trillion parameter follow-up

On July 19, Ali Tongyi released Qwen3.8-Max-Preview, with a parameter scale of 2.4 trillion yuan. The three major channels of Token Plan, Qoder, and QoderWork will be launched for the first time. The official version will be launched and open source in the near future. Following the Kimi K3 (2.8T) and DeepSeek V4 (1.6T), this is another domestic flagship with a 1.5T class or higher. Alibaba uses a “cloud+model+programming tool chain” to monetize in a closed loop, and ecological stickiness continues to increase.

Risk warning: model iteration and commercialization falling short of expectations; computing power, production capacity and supply chain risks; regulatory policy changes; macro and geopolitical risks.