GF Securities: Domestic model capabilities have improved markedly, focusing on the easing of CSP computing power constraints and the pace of order growth
The Zhitong Finance App learned that Guangfa Securities released a research report saying that the core drive of this round of the AI industry chain comes from enhanced model agentic coding capabilities. Programming tools represented by Claude Code and Codex are shifting from auxiliary completion to autonomous collaboration, and this is spreading multi-level investment opportunities upstream and downstream. Looking upward, Agentic Workload's consumption of computing power has increased dramatically, driving high growth in cloud vendor IaaS & MaaS business, reducing market concerns about issues such as AI bubbles caused by overinvestment by cloud vendors. Sensitivity to the decline in FCFmargin has declined, and attention is being paid to the easing of CSP computing power constraints and the pace of order growth. Looking at the model makers themselves, Agentic tasks drive a sharp rise in token consumption and an accelerated increase in the ARR of the leading model. Short-term transactions focus on marginal changes such as model updates, ARR data, token prices, and computing power supply. Looking at downstream applications, Coding Agent was the first to verify PMF, focusing on the expansion space brought about by “Claude Code for X” style scenario replication and improved token economy.
The main views of GF Securities are as follows:
AI topics
The models were compared horizontally, and the capabilities of domestic models improved markedly. Recently, a number of domestic and foreign AI manufacturers have successively launched a new generation of flagship models. Following Anthropic's release of Claude Fable 5 and the release of GPT-5.6 by OpenAI, Dark Side of the Moon released the Kimi K3 model on July 16. We believe that the leading flagship models have all crossed the threshold of being usable in investment and research scenarios. According to our tests, the functionality of all four models has been fully met. The gap between the models has changed from “whether it can be completed” to “completion quality and information density”, and the task capability of the domestic model has improved significantly, further narrowing the gap with the leading model.
AI data tracking
Recently, the number of token calls through the OpenRouter large model has increased. According to weekly data, the number of large model tokens called via OpenRouter on July 6, 2026 and July 12, 2026 was 52.6 tons. Compared with 46.7T on June 29, 2026 and July 5, 2026, the number of subsequent model calls is expected to continue to rise. Among the models called through OpenRouter, Hy3 (free), which was newly launched by Tencent on July 6, promoted call volume to the top of the list with a free open strategy. MiMoV2.5 maintained second place, and DeepSeek V4 Flash ranked third.
AI hotspot tracking
Dark Side of the Moon released Kimi K3 and announced the simultaneous launch of the model Kimi, Kimi Work, Kimi Code, and KimiAPI. Kimi K3 is a 2.8 trillion parameter MoE model, built on Kimi Delta Attention (KDA) and attentionResiduals (attnRes) architectures, and natively supports visual understanding and a 1 million token context window. Officials said that KiMiK3 is the first open source 3-trillion parameter model (2.8 trillion parameter scale). The full model weights will be released in open source before July 27, and the technical report will be released simultaneously.
Risk warning: Model iteration results fall short of expectations; commercial application implementation falls short of expectations; risks of copyright, ethics, and content quality; risk of incomplete evaluation due to limited sample size and access rights.