人工智能周报（26 年第4 周）：MiniMax Agent 2.0 正式发布，百度文心 5.0 上线

Investment Rating - The report maintains an "Outperform" rating for the industry, indicating expected performance above the market benchmark by over 10% [3][28]. Core Insights - The report anticipates a surge in mature AI agent products in 2026, driven by advancements in multi-modal capabilities, long text processing, and reasoning abilities. This increase in demand for reasoning will boost revenues for upstream cloud computing providers [2][25]. - Domestic internet giants are approximately one year behind their overseas counterparts in AI capital expenditures. As the capabilities of large models improve and supply builds up, AI will increasingly empower the core businesses of these giants [2][25]. - The third quarter is expected to be a peak for investment in the internet giants' food delivery competition, with a projected narrowing of losses for Alibaba, Meituan, and JD.com in the fourth quarter [2][25]. - The report recommends focusing on AI-related stocks, specifically highlighting Alibaba and Tencent Holdings as key investment opportunities [2][25]. Company Dynamics - ByteDance launched version 2.0 of its AI agent platform "Coze," introducing new features such as Agent Skills and Agent Plan, allowing users to set long-term goals for AI to manage [17]. - Anker and Feishu jointly released the "AI Recording Bean," a portable AI hardware device designed for various recording scenarios [17]. - MiniMax's AI native workspace Agent 2.0 was officially launched, featuring components that enhance task execution and business understanding [19]. - The American AI startup Humans& secured $480 million in seed funding, achieving a valuation of $4.48 billion [19]. - Tesla's humanoid robot Optimus is set for public sale by the end of 2027, with a target price of $20,000 [20]. - Google Gemini introduced a free SAT simulation feature in collaboration with The Princeton Review, providing instant feedback to users [20]. - xAI Grok Imagine launched a 10-second video generation feature, enhancing its capabilities in the AI video sector [21]. Underlying Technology - Zhipu AI released and open-sourced the GLM-4.7-Flash model, a lightweight large language model designed for local programming and intelligent assistance [22]. - DeepSeek unveiled a new model architecture called "MODEL1," which is expected to be efficient for inference tasks [22]. - Alibaba's Tongyi Qianwen open-sourced the Qwen3-TTS series voice generation model, supporting multiple languages and dialects [23]. - Baidu launched the official version of its Wenxin model 5.0, which boasts a parameter scale of 24 trillion and excels in multi-modal understanding and generation [23]. - Google DeepMind introduced the D4RT model, significantly improving the speed of dynamic 4D reconstruction [24].