直击WAIC2025｜手机Agent竞赛升级：荣耀发布多模态感知大模型MagicGUI，从单智能体任务执行到多智能体协同

Core Insights - The article emphasizes that the era of AI in smartphones should extend beyond basic functionalities like translation and document processing, advocating for a broader imagination of AI's capabilities in mobile devices [1] - Honor's release of the MagicGUI model, with 7 billion parameters, marks a significant advancement in AI assistants, evolving from traditional voice assistants to more capable digital assistants that can understand complex needs and execute multi-step tasks [1][2] Group 1: Evolution of AI Assistants - Since the rise of large models in 2023, major smartphone manufacturers have recognized the shift from basic voice assistants to lightweight intelligent agents capable of perception, reasoning, decision-making, and operation [2] - Honor's YOYO has evolved from executing single tasks to coordinating multiple intelligent agents, showcasing a significant leap in functionality compared to traditional voice assistants [2][7] Group 2: Comparison with Competitors - Apple's Siri, introduced in 2011, has seen limited updates and remains largely underutilized, while Android counterparts like Honor's YOYO, Vivo's "Blue Heart Little V," and Xiaomi's "Super Xiao Ai" have advanced to task-oriented intelligent agents capable of executing complex tasks [5][6] - The transition from app-driven interactions to agent-driven frameworks signifies a major shift in user-device interaction, with AI assistants taking the lead in understanding and executing tasks [8] Group 3: Technical Advancements - The MagicGUI model employs a two-phase training paradigm, enhancing the model's screen perception and positioning capabilities through large-scale GUI knowledge injection and reinforcement learning [9] - The trained MagicGUI model allows YOYO to think and act based on visual information from the screen, improving its efficiency and adaptability in task execution [9] Group 4: Open Source Initiative - Honor has announced that the MagicGUI model and related testing data will be made available on open-source platforms, promoting collaboration and further development in the field [9]