GF SEC: Significant progress in domestic model capability, focusing on alleviating CSP computation constraints and the pace of order growth.

GF SEC released a research report stating that the core driving force of this round of AI industry chain comes from the enhancement of model agentic coding capabilities. Programming tools represented by Claude Code and Codex are transitioning from auxiliary completion to autonomous collaboration, leading to multiple layers of investment opportunities upstream and downstream. Looking upstream, Agentic Workload has significantly increased the consumption of computing power, driving high growth in cloud vendors' IaaS & MaaS businesses, reducing concerns about market oversaturation leading to AI bubbles, and decreasing sensitivity to FCF Margin decline. Focus on CSP computing power constraints relief and order growth pace. Looking at the model vendors themselves, agentic task-driven Token consumption has surged, and top model ARR has accelerated growth. In the short term, the focus is on model updates, ARR data, Token prices, and computing power supply changes. Looking downstream at applications, Coding Agent is the first to validate PMF, focusing on the expansion space brought about by the scene replication of "Claude Code for X" and the improvement in Token economics. GF SEC's main points are as follows: AI Topic Horizontal comparison of models, significant progress in domestic model capabilities. In recent times, many domestic and foreign AI vendors have successively released new flagship models. After Anthropic released Claude Fable 5 and made it available, OpenAI released GPT-5.6, and Yuezhi An Mian released Kimi K3 on July 16th. We believe that the flagship models have all crossed the threshold of usability in investment research scenarios. Our tests show that the functionality of the four models meets the standards, and the gap between the models has shifted from "can it be completed" to "quality of completion and information density." The task capabilities of domestic models have significantly improved, further narrowing the gap with the top models. AI Data Tracking Recently, an increase in token calls for large models through OpenRouter. Weekly data shows that the number of large model tokens called through OpenRouter from July 6th, 2026, to July 12th, 2026, was 52.6T, an increase from 46.7T from June 29th to July 5th, 2026. It is expected that the subsequent model call volume will continue to increase. Among the models called through OpenRouter, Tencent's newly launched Hy3 (free) on July 6th drove the call volume to the top of the list with a free open strategy, while MiMoV2.5 maintained the second position, and DeepSeek V4 Flash ranked third. AI Hotspot Tracking Yuezhi An Mian released Kimi K3 and announced the simultaneous launch of Kimi, Kimi Work, Kimi Code, and KimiAPI models. Kimi K3 is a 2.8 trillion parameter MoE model, built on the Kimi Delta Attention (KDA) and AttentionResiduals (AttnRes) architecture, natively supporting visual understanding and a 1 million token context window. The official stated that KimiK3 is the first open-source 3 trillion parameter model (2.8 trillion parameters in scale), and the complete model weights will be released open-source by July 27th, along with the publication of a technical report. Risk Warning: Model iteration effects may not meet expectations; commercial applications may not materialize as expected; risks related to copyright, ethics, and content quality; risks of limited sample size and access restrictions leading to incomplete evaluation.