Everbright Securities: Maintaining the Xiaomi Group-W (01810) “buy” rating Xiaomi's AI strategy is an endurance race rather than a single point sprint

The Zhitong Finance App learned that Everbright Securities released a research report saying that on December 17, Xiaomi Group-W (01810) held the 2025 “Full Ecosystem of People and Cars” partner conference. The core highlights focused on breakthroughs in big model technology, deepening collaboration across the entire ecosystem, and open ecological expansion. Xiaomi's AI strategy is long distance running rather than sprinting, medium- to long-term. I am optimistic about the overall layout of the large model of Xiaomi's self-developed chip+Surge OS+AI base, and the “human-car” ecological collaboration will provide the company with a scenario for large-scale AI applications in the future. Maintain the company's “buy” rating.

The main views of Everbright Securities are as follows:

Mimo-V2-Flash model: excellent speed and cost, looking forward to future hardware integration and commercialization progress

Xiaomi released the next-generation base model Mimo-V2-Flash. 1) In terms of performance, the model has a lightweight design with “309B total parameters+15B activation parameters”, ranking in the top 2 global open source model evaluations for code capability and agent capability. The core evaluation indicators surpass DeepSeek V3.2 and Kimi K2-Thinking, which have 2 to 3 times the number of parameters, leading the industry in technical strength. 2) In terms of technical architecture, Mimo-V2-Flash uses a 5:1 hybrid SWA (sliding window attention+full attention) architecture, fixed KVCache adapts to mainstream inference frameworks, and is equipped with 3-layer MTP inference optimization to achieve 2.0 to 2.6 times faster inference, a stand-alone throughput of 5000-15,000 tokens/s, and a single request throughput of 115-150 tokens/s. 3) In terms of inference efficiency, Mimo-V2-Flash has significant cost and speed advantages. Its inference speed is 3 times that of DeepSeek V3.2, which has a similar inference cost. The inference cost is 1/20 of Gemini 2.5 Pro with similar inference speed. It can effectively solve the efficiency and cost pain points of large-scale model implementation, and provide core technical support for Xiaomi's “human-car” real-time interaction in all scenarios. 4) At present, the model is fully open source, and the API and experience web are simultaneously open, which is expected to attract global developers to accumulate application cases, feed back model iteration, and eventually form a virtuous cycle of “Xiaomi model - developer - end user”.

Hardware ecosystem: basing the scale+breaking the boundaries of collaboration, building a solid foundation for AI implementation

At present, Xiaomi has built the world's largest consumer-grade hardware platform, with 1.8 billion hardware connected devices worldwide, including 740 million monthly active mobile phone users, more than 1 billion IoT connected devices, delivered more than 500,000 cars, and CariOT has opened more than 30 types of rear-mounted components, forming a full-scenario hardware coverage of “personal devices+travel devices+home devices”. The bank believes that the scale advantage of Xiaomi's hardware terminals and multi-category collaboration is expected to provide broad implementation scenarios and data sources for its AI technology.

Software system: technical base+open source ecosystem, opening up AI collaboration links

1) Surge OS 3: Released in August 2025, as the core foundation for ecological collaboration, achieving four major capability upgrades. ① In terms of full-ecosystem applications, the Super Island function is connected to 60+ apps, adapts to 70+ scenarios, and has been adapted to multiple terminal displays, and has been adapted to 128 million app installations; ② In terms of interconnection capabilities, Xiaomi has built a huge connected ecosystem with daily activity of 410 million, with an average daily interconnection of over 270 million interconnect operations, achieving senseless interconnection across the Apple and Windows ecosystems, and supporting immediate functions such as cloud album sync and cross-device unlocking; ③ In terms of intelligent experience, the AI penetration rate of Xiaomi's system-level functions continues to increase, reaching 120 million monthly active users on mobile phones and tablets, and more than 120 million daily active users. 65 million times. As the core AI portal, Super Xiaoai has the ability to achieve four major upgrades: Q&A, scene perception, screen circle, and direct access to four major upgrades; ④ In terms of security and privacy, the next-generation permission system simplifies developer access processes, enables unified adaptation across Android manufacturers, and provides security guarantees for AI applications. 2) Auxiliary software ecosystem: As a powerful OS subsystem, the Vela lightweight system has empowered 160 million smart devices, and has surpassed 100 open source ecosystem partners to achieve key expansion from IoT chips to automotive MCUs, supporting multi-terminal lightweight AI deployment; the MINT platform acts as an end-to-end AI development pipeline, integrating 3.9T data and 20+ mature algorithms to help developers quickly complete high-complexity AI deployments. It has been verified in products such as AI cat eye for door locks and AI obstacle avoidance for sweeping robots.

Profit forecasting and valuation

At this conference, Xiaomi emphasized the R&D investment plan of over 200 billion yuan over the next five years (of which 40 billion yuan is expected in 2026). In terms of profit forecasting, the bank maintains the 25-27 non-IFRS net profit forecast of 426/438/51 billion yuan.

Risk warning: mobile phone shipments are under pressure, IoT gross margin falls short of expectations, AI expansion falls short of expectations, etc.