Zheshang Securities has released a research report stating that the large model industry is increasingly focusing on inference and multimodal approaches in its technological paradigm. The current phase places greater emphasis on the feasibility of commercial deployment at the application layer and improving the return on investment efficiency of technological expenditure. Among these developments, the new Qwen model achieves superior performance with a smaller size, significantly boosting inference efficiency. At the application layer, it is embedded within Alibaba's internal ecosystem to enhance customer acquisition, retention, and data feedback loops, signaling a shift for AI from conversational chat to performing practical tasks. Competition for consumer-facing entry points among AI applications intensified during the Spring Festival period, with Qwen's Daily Active Users showing significant growth and accelerating its catch-up pace. The main viewpoints of Zheshang Securities are as follows:

**Industry Trends for Large Models: Paradigm Shift Towards Inference and Multimodal, Emphasizing Application Deployment and ROI** Since 2025, large model development has exhibited the following trends: 1) Capability evolution from "generation" to "reasoning," 2) A paradigm emphasizing native multimodality, 3) Increased focus on commercial scenario deployment and ROI, and 4) Gradual enhancement of large models' agentic capabilities. On February 20th, Zhipu and MiniMax saw significant stock price increases, closing up 42% and 14% respectively. This was primarily driven by Zhipu's release of its new flagship model, GLM-5, on February 12th, which achieved state-of-the-art open-source results in programming and agent capabilities. User experience in real programming scenarios is reported to be approaching that of the currently recognized top-tier closed-source model, Claude Opus 4.5. Additionally, Zhipu announced a 30% overall price increase for its GLM Coding Plan套餐, potentially indicating a new phase for domestic large models shifting from traffic consumption to monetizing computational value. Regarding investment targets, it is suggested to monitor: Zhipu (02513) and MINIMAX-WP (00100).

**Qwen's New Model "Punches Above Its Weight," Fully Integrates with Alibaba Ecosystem, Shifting AI from Chat to Task Execution** Qwen's technological evolution places greater importance on efficiency over simply scaling parameters, emphasizing multimodal training including vision. Its agent capabilities are shifting focus from dialogue towards ecosystem synergy and practical implementation, with a cost structure more focused on ROI. It aims to build core competitiveness through full-stack coordination of model-chip-cloud. On February 16th, Alibaba open-sourced its new-generation large model Qwen3.5-Plus. This model achieves strong performance with a smaller footprint, having a total of 397 billion parameters but only activating 17 billion, reducing GPU memory usage for deployment by 60% and significantly boosting inference efficiency, with maximum inference throughput increasing by up to 19 times. Its API price is as low as 0.8 RMB per million tokens. Qwen is now integrated with businesses including Taobao Flash Sales, Taobao, Alipay, Fliggy, Amap, and Damai, enabling functions like ordering takeout, selecting products, booking flights/hotels, choosing restaurants, and buying movie tickets with a single command. Zheshang Securities believes that Qwen, as a new consumer traffic entry point fully connecting the lifestyle consumption chain for吃喝玩乐 (food, drink, entertainment, fun), could bring additional traffic and order volume to downstream service providers. Recommended target: Alibaba-W (09988). Suggested to monitor: CaoCao Chuxing (02643), SF Intra-City (09699), Damai Entertainment (01060), Focus Media (002027.SZ), Bilibili-W (09626).

**AI Application Consumer Positioning: Qwen's DAU Shows Notable Increase During Spring Festival Red Envelope Activities, Catch-Up Accelerates** During the Spring Festival period, various AI applications used red envelope campaigns as an entry point, leveraging their respective platform ecosystems to compete for consumer traffic. According to Questmobile, on February 7th, Qwen's DAU reached 73.52 million, while Doubao's DAU was 78.71 million; according to Tencent, Yuanbao's DAU had exceeded 50 million by February 18th. In terms of campaign scale, Alibaba's Qwen led with 3 billion RMB and launched its new-generation model, "Qwen Qwen3.5-Plus," during the holiday, achieving a generational leap from pure text to native multimodal capabilities. Doubao holds a first-mover advantage and also leverages hardware for integrated software-hardware synergy; Alibaba's Qwen primarily synergizes with its group's internal ecosystem, already integrated with Taobao Flash Sales, Fliggy, Amap, and others; Tencent's Yuanbao mainly relies on WeChat's "social fission" for growth. The AI red envelope battle represents an attempt to shift from "buying traffic" to "buying user mindshare" and "buying user pathways" to achieve a closed loop for customer acquisition and retention. Recommended target: Tencent Holdings (00700).

**Risk Warning:** Technology iteration may fall short of expectations; potential safety and compliance risks.