Caitong: Continues to be bullish on on-device AI; attention should be paid to opportunities brought by upgrades in personal agent capabilities and other factors.

Caitong: Continues to be bullish on on-device AI; attention should be paid to opportunities brought by upgrades in personal agent capabilities and other factors. Caitong released a research report stating that as the market's main theme gradually spreads from cloud training to terminal applications and entry-point competition, on-device AI has the conditions to be repeatedly traded in a weak market environment. The dominant variable in this round of industrial evolution is shifting from hardware specification-driven to model and personal Agent capability-driven. Hardware is the physical extension of Agents, and the combination of XR and world models is expected to open up the imagination space for the next stage of on-device AI. The bank continues to be bullish on on-device AI and attaches importance to opportunities brought by upgrades in personal agent capabilities, the launch of new hardware, and improvements in XR content supply driven by world models. Caitong's main views are as follows: The on-device AI direction has the foundation for independent trading in a weak market environment, because its catalysts come from product and industry progress rather than a comprehensive improvement in aggregate liquidity. At the A-share level, the market has recently been in an atypical tailwind environment. However, the current industrial catalysts for on-device AI come from terminal products, Agent releases, regulatory filing implementation, and penetration rate increases, belonging to a typical "bottom-up product innovation-driven" rather than purely macro-liquidity-driven pattern. As the market's main theme gradually spreads from cloud training to terminal applications and entry-point competition, on-device AI has the conditions to be repeatedly traded in a weak market environment. On-device AI is moving from feature upgrades to product-definition reconstruction, and the dominant variable in this round of industrial evolution is shifting from hardware specification-driven to model and personal Agent capability-driven. Unlike previous rounds of on-device market activity, which mostly centered on terminal parameter upgrades and enhancements of partial features, the reason current AI hardware can continue to be launched is essentially that model capabilitiesespecially memory, planning, tool calling, multimodal understanding, and continuous task executionhave improved markedly, and personal Agents are beginning to have stronger usability. This means the core of industrial competition is no longer just hardware specifications themselves, but model efficiency, system scheduling, tool calling, cross-device collaboration, and task closed-loop capabilities; the importance of hardware has not declined, but its role is increasingly shifting toward an entry point and carrier that undertakes and amplifies model capabilities. Hardware is the physical extension of Agents, and the combination of XR and world models is expected to open up the imagination space for the next stage of on-device AI. The true value of on-device hardware lies in continuously acquiring visual, auditory, and contextual information about the user's environment through multimodal sensors such as cameras, microphones, and location and spatial perception, thereby helping Agents understand scenarios more completely and improving task execution efficiency and productivity value. In terms of product cadence, Meta has clearly advanced Muse Charm to ship in December 2026 and plans to integrate Muse into smart glasses around November this year; on the OpenAI side, CodexMicro will go on sale in July 2026, and according to Cailianshe, it is expected to successively release on-device AI products such as brooches and recording devices around late 2026 or early 2027. On the XR side, world models are driving the industry from display devices to contextual interaction terminals, and are improving the consistency, real-time nature, and physical interactability of virtual environments. For XR, world models are expected to become the underlying support that complements content, interaction, and spatial understanding capabilities, driving XR to evolve from display terminals to Agent perception and action interfaces. Risk warnings: personal agent capabilities and application implementation fall short of expectations, privacy authorization and data compliance restrict usage scenarios, new product launches and market acceptance fall short of expectations, improvements in hardware performance, cost, and manufacturing yield fall short of expectations, conversion of world models into XR content falls short of expectations, and intensifying industry competition suppresses profitability.