XIAOMI(HK:01810) 光大证券研究·2025-12-20 00:04Core Viewpoint - Xiaomi Group is focusing on breakthroughs in large model technology, deepening ecosystem collaboration, and expanding an open ecosystem as highlighted in their 2025 "Human-Vehicle-Home Ecosystem" partner conference [4]. Group 1: Large Model Technology - Xiaomi has launched the new base large model Mimo-V2-Flash, which features a lightweight design with "309 billion total parameters + 15 billion activation parameters," ranking in the top 2 globally in code capability and agent capability evaluations [5]. - The Mimo-V2-Flash model employs a Hybrid SWA architecture with a 5:1 ratio, achieving a 2.0 to 2.6 times inference speedup, with single-machine throughput of 5000-15000 tokens per second and single-request throughput of 115-150 tokens per second [5]. - The model's inference speed is three times that of DeepSeek V3.2, while its inference cost is 1/20 of Gemini 2.5 Pro, addressing efficiency and cost challenges in large model deployment [5]. Group 2: Hardware Ecosystem - Xiaomi has established the world's largest consumer hardware platform, with 1.8 billion connected devices globally, including 740 million monthly active mobile users, over 1 billion IoT devices, and more than 500,000 cars delivered [7]. - The integration of personal, travel, and home devices creates a comprehensive hardware coverage that provides ample scenarios and data sources for AI technology deployment [7]. Group 3: Software Ecosystem - The upcoming release of the Surge OS 3 in August 2025 will enhance ecosystem collaboration with upgrades in application integration, interconnectivity, smart experience, and security [8]. - Xiaomi's ecosystem supports over 60 applications and 70 scenarios, with 1.28 billion application installations, and a daily active interconnect ecosystem of 410 million users [8]. - The Vela lightweight system has empowered 160 million smart devices, and the MINT platform integrates 3.9 terabytes of data and over 20 mature algorithms to facilitate complex AI deployments [8].