据了解，千问App接入淘宝、支付宝、淘宝闪购等阿里生态业务，可以实现点外卖、买东西等AI购物功能。此前，豆包团队也曾进行类似 的"试水"并引发热议。 AI时代的到来似乎也同时让部分APP开始"退居幕后"，受访专家指出，它意味着我们正从"以应用为中心"的APP时代，跃迁至"以用户意图为中 心"的智能体时代。用户只需用最自然的方式表达需求，AI智能体便能自动拆解任务、调度后台资源、完成任务闭环。这将破解我国数字经济 发展所面临的流量瓶颈，不仅提升单个用户的任务完成效率，更在宏观层面促进整个社会商业活动的流转速度和资源配置效率。 与此同时，从数字经济时代到智能经济新时代，竞争逻辑从"基于产品与服务的市场竞争"向"基于'人工智能+'的生态竞争"的转变也正浮出水 面。 破解数字经济流量瓶颈 根据阿里巴巴发布的内容，基于Qwen最强模型与阿里最丰富生态，千问App接入淘宝、支付宝、淘宝闪购、飞猪、高德等阿里生态业务，在 全球首次实现点外卖、买东西、订机票、订酒店等AI购物功能，向所有用户开放测试。 21世纪经济报道记者冉黎黎 北京报道 在"传统APP可能沦为智能体背后的被感知被调度资源"等一系列预言之下，2026年1月1 ...

来源：爱分析ifenxi 序言 在中国企业级应用服务发展的数十年历程中，云大厂、模型大厂与垂直应用厂商之间曾维持着长期的生态平衡，双方以基础设施、业务应用为界，各司其 职。然而，随着智能体时代的到来，这一传统边界正面临前所未有的冲击与重构。过去，由于大厂缺乏对knowhow的理解，其通用能力在深水区往往无所适 从；但在智能体逻辑下，企业需求正从流程管理转向结果交付，基础模型对原始知识的直接利用能力也显著增强，这使得大厂得以跨越藩篱，直接切入应用 层的核心价值地带。 本报告的核心受众为大厂及应用厂商的中高层决策者。报告旨在通过剖析智能体时代的越界逻辑，解决双方在业务变迁过程中的边界问题。我们将通过建立 基于任务复杂度与知识复杂度的象限判定模型，提供一张清晰的风险地图，用以判断哪些应用处于大厂的延长线上，哪些应用仍能坚守护城河。 此外，本报告不仅是一份风险评估，更是一份行动指南。针对处于不同象限的应用场景，我们为应用厂商提供了突围策略；同时也为大厂建议了从直接进攻 到生态赋能的进击边界。通过本报告的分析，我们期望推动大厂与应用厂商在新的智能体生态体系中实现利益绑定与能力互补，共同构筑智能化时代的产业 新轮廓。 第 ...

Core Insights - The core theme of the China Wealth Management 50 Forum 2025 Annual Meeting is "Towards the Construction of a Financial Power during the 14th Five-Year Plan" [1] AI Application in Finance - AI applications in the financial industry are transitioning from the "usable" stage to the "useful" stage, with 90% of success depending on engineering capabilities [3][8] - The development of large models is entering the "Agentic" era, where AI will autonomously complete tasks and create business value [3][8] - AI will lead to structural changes in financial institutions, reshaping competitive barriers in five key areas: 1. Redefining traffic entry points from passive app clicks to proactive intent recognition through intelligent services 2. Redefining financial products and services for deep customization 3. Restructuring the entire user journey to make financial services more inclusive 4. Redefining operational objects and forms, with intelligent assistants becoming the main channel influencing customer mindset 5. Ultimately affecting talent and organization, moving towards a "human-machine coexistence" state [3][8] Huawei's AI Strategy - Huawei's financial AI strategy aims to support the industry in moving from "usable" to "useful," providing a full-stack capability from advanced Ascend computing power to a one-stop AI development platform (ModelArts) and an intelligent agent development and operation platform (Versatile Agent) [3][8] - The strategy includes talent training courses and focuses on three typical scenarios co-created with leading financial institutions [3][8] Specific Use Cases - In mobile banking app scenarios, Huawei uses models like Pangu 7B to enhance service accuracy to over 95% while optimizing computing power and reducing costs, achieving end-to-end latency under 2 seconds [4][9] - In intelligent risk control scenarios, the core solution involves converting expert experience into "thinking chain" data and using large models with "slow thinking" capabilities for reinforcement learning, ensuring real-time updates and high accuracy of risk control models [4][9] - For report generation (applicable to credit and investment research), an innovative "Deep Research" development paradigm allows intelligent agents to automatically organize tasks and generate high-quality reports through repeated interactions with external data sources and knowledge bases [4][10] Engineering Challenges and Recommendations - The financial industry, characterized by strong regulation and high standards, faces challenges in engineering rather than merely applying generic models or external knowledge bases [5][10] - To address systemic latency, accuracy, humanization, and cost issues, strong dynamic business orchestration capabilities are required, along with complex model tuning, intelligent agent tuning, system integration, and full-link monitoring [5][10] - Eight recommendations for financial institutions include: 1. AI should be a company-level strategy led by top management 2. Business departments must deeply participate in building integrated teams of technology, business, and data 3. Focus on "useful" applications rather than "showcase" applications, paying attention to actual usage metrics 4. Adopt diversified models and open architectures 5. Combine engineering experience from professional fields 6. Build enterprise-level AI pipelines and regulatory-compliant governance systems 7. Develop high-quality datasets 8. Recognize that 90% of success depends on engineering capabilities [6][10]

Core Insights - Google has launched Gemini 3 Flash, which replaces Gemini 2.5 Flash as the default model in the Gemini application, allowing millions of users to access its capabilities for free [1] - Gemini 3 Flash offers high performance at a significantly lower cost, achieving a score of 78% in the SWE-bench Verified benchmark, surpassing both Gemini 2.5 and even outperforming Gemini 3 Pro in certain areas [1][3] - The model is designed for high-frequency, rapid development scenarios, enabling real-time application updates and simplifying workflows for users [2][6] Pricing and Cost Efficiency - The pricing for Gemini 3 Flash is set at $0.50 per million input tokens and $3.00 per million output tokens, making it highly competitive compared to previous models [2][5] - The introduction of Gemini 3 Flash significantly lowers the entry barrier for advanced AI capabilities, which were previously costly, thus fostering a more accessible AI environment [5] Performance and Benchmarking - Gemini 3 Flash demonstrates superior performance across various benchmarks, achieving three times the speed of Gemini 2.5 Pro and excelling in tasks requiring high precision and rapid feedback [5][9] - In specific tests, Gemini 3 Flash outperformed flagship models like GPT and Claude in certain dimensions, indicating its strong capabilities in automation and real-time processing [3][4] Applications and Use Cases - The model is being integrated into various sectors, including software engineering, legal, and financial industries, where it enhances response times and accuracy in complex tasks [9][11] - Gemini 3 Flash's multi-modal capabilities allow for rapid transformation of unstructured data into actionable insights, proving its utility in diverse applications from legal document analysis to game development [6][11] Ecosystem and Market Impact - The launch of Gemini 3 Flash signifies a strategic enhancement of Google's AI ecosystem, positioning it as a leader in the competitive landscape of AI models [9][10] - The model's capabilities are expected to drive widespread adoption of AI technologies across industries, marking a shift towards more integrated and efficient AI solutions [8][13]

Core Insights - The article highlights the transition of artificial intelligence (AI) into the "intelligent agent era" by 2025, marking a shift from tool-based products to autonomous decision-making systems [1] - China's AI sector is evolving from "made in China" to "created in China," positioning itself as a leader in the global AI race with unique paths in performance, cost, and algorithm originality [1] - The Yangtze River Delta region is emerging as a hub for AI innovation, with a surge of new AI products and applications in recent months [1] Technological Breakthroughs - The Yangtze River Delta's AI models are establishing a robust foundation for intelligent agents, exemplified by MiniMax's M2 model, which ranks among the top five in global assessments with only 10 billion activation parameters [2] - MiniMax's technology has been integrated by global tech giant Meta, marking a significant recognition of Chinese AI algorithms on an international scale [2] - MiniMax has expanded its offerings with a comprehensive suite of models across text, video, voice, and music, showcasing a breakthrough in multimodal AI capabilities [2] Vertical Industry Advantages - Companies in the Yangtze River Delta are demonstrating specialized strengths, such as Nanjing's NanZhi Optoelectronics, which upgraded its photon-specific model to enhance design efficiency by 30% [3] - The "Inspiration Intelligent Agent" developed by Hefei's team revolutionizes visual content creation, allowing users to perform complex tasks through simple dialogue interactions [3] - The team received the Best Demonstration Award at the ACM International Multimedia Conference, highlighting their innovative contributions to the field [3] Scene Empowerment - The rich industrial ecosystem and large consumer market in the Yangtze River Delta facilitate the practical application of AI models, as evidenced by the rapid success of Ant Group's "Lingguang" application, which surpassed 2 million downloads within a week [4][5] - The "Qwen" model from Alibaba also achieved over 10 million downloads shortly after its public testing, showcasing the competitive edge of Chinese AI products [5] - Suzhou's industrial park has seen significant algorithm aggregation, with 35 algorithms approved for deep synthesis services, leading the province in this area [5] Policy Support - The article discusses the comprehensive "ecological cultivation" system established in the Yangtze River Delta, supported by forward-looking policies from key cities [7] - Shanghai has positioned AI as one of its three leading industries, with projections indicating the industry will exceed 450 billion yuan by 2025 [7] - Hangzhou aims to create a full-chain support system for AI innovation, with the launch of the largest AI open-source community in China [7] City-Specific Initiatives - Suzhou's action plan aims for over 3,000 AI companies and a core industry scale growth of over 20% annually by 2026 [8] - Nanjing plans to achieve a core industry scale of 60 billion yuan by 2026, focusing on developing foundational and industry-specific AI models [8] - Hefei is advancing its AI industry through a dual strategy, establishing a comprehensive ecosystem that includes data labeling and computational support [9]

Investment Rating - The report indicates a strong growth potential for the AI Agent industry, with a projected market size exceeding 35.7 billion yuan by 2029, reflecting a compound annual growth rate (CAGR) of 52.4% [8][9]. Core Insights - The AI Agent industry is positioned as a key commercial application of large models, showcasing significant potential for market expansion and value creation [8][9]. - The growth of the large model market is driven by three main factors: optimization of computing power and infrastructure, exponential growth in data resources, and the increasing demand for digital transformation across industries [12][16]. - AI large models are reshaping enterprise value systems through internal process integration and external product innovation, leading to a comprehensive transformation of business operations and models [18]. - The commercialization of AI large models is evolving in three layers: embedded applications, native applications, and hardware applications, with embedded applications being the most mature [22][23]. Summary by Sections AI Large Model Market Size and Growth Forecast - By 2029, the Chinese large model market is expected to exceed 140 billion yuan, driven by overall infrastructure expansion [8][9]. Growth Drivers of the Large Model Market - The expansion of the large model market is propelled by advancements in computing power, improvements in data quality and governance, and a surge in digitalization and intelligent upgrade demands across various sectors [12][16]. AI Large Model Empowering Enterprise Value Reconstruction - AI large models enhance operational efficiency and user experience, leading to significant value creation for enterprises [18][19]. Commercialization Development Status of AI Large Models - The commercialization of AI large models is characterized by a three-tier evolution, with embedded applications being the most developed, while native applications and hardware applications are still in exploratory and early stages, respectively [22][23]. AI Large Model Product Usage Flow Distribution - The flow distribution of AI large model products shows that dialogue assistants dominate both web and mobile platforms, indicating a concentrated user engagement [25][26]. AI Agent Product System - The AI Agent product system consists of three layers: general-purpose, business-specific, and industry-specific, facilitating rapid coverage and validation across various sectors [27][28]. AI Agent Supply Scene Distribution - The supply side of AI Agent products is primarily focused on general scenarios, accounting for nearly 70% of the market, leveraging technological versatility and cost advantages [28][29]. AI Agent Industry Demand Scene Distribution - Demand for AI Agents is concentrated in high-frequency interaction scenarios such as e-commerce, finance, and education, with advanced manufacturing also experiencing growth through digital transformation [32].

Core Insights - John Sculley, former CEO of Apple, stated that OpenAI has become Apple's first real competitor in decades, emphasizing that artificial intelligence is not a particularly strong area for Apple [1][3] Group 1: Apple's Position in AI - Apple's performance in the AI race is lagging compared to competitors like OpenAI, Google, Amazon, and Meta, which are continuously updating their products [3] - Apple's plans to upgrade its AI assistant Siri faced delays earlier this year, marking a significant setback in product launches [3] Group 2: Future Leadership and Business Model Shift - Speculation surrounds the potential retirement of current CEO Tim Cook, with Sculley suggesting that whoever succeeds him must lead Apple from an application-centric era to an agent-centric era [3] - In the agent-centric era, intelligent agents will replace many applications and autonomously complete complex tasks, posing a significant challenge to Apple's existing business model [3] - Sculley believes that AI-driven intelligent agents will help knowledge workers automate cumbersome workflows, prompting more tech companies to shift towards subscription-based business models, which he views as more advantageous than the current application-centered model [3] Group 3: Collaboration with OpenAI - Notably, former Apple design chief Jony Ive recently appeared at OpenAI, where the company acquired his device startup for over $6 billion earlier this year [4] - Ive aims to develop devices that address issues arising from smartphones and tablets since their inception, and Sculley recognizes his capabilities, suggesting that his collaboration with OpenAI CEO Sam Altman could lead to breakthroughs in the field of large language models [4]

Core Insights - The article emphasizes that the capabilities of Li Auto's MindGPT 3.1 are significantly underestimated, highlighting three main anchors of value [1] - MindGPT 3.1's ASPO incorporates innovative optimizations from DeepSeek R1's GRPO, showcasing Li Auto's ability to rapidly learn and internalize the best practices in AI [1][8] - There is a lack of in-depth discussion about Li Auto's technology in the information ecosystem, indicating a potential undervaluation of its advancements [1] Performance Metrics - MindGPT 3.1 is a fast reasoning language model, achieving speeds of up to 200 tokens per second, nearly five times faster than MindGPT 3.0, which is a significant improvement compared to GPT-4's maximum of 79.87 tokens per second [2][4] - The model shows notable enhancements in tool invocation accuracy, complex task completion rates, and response richness compared to its predecessor [4] Benchmarking Results - MindGPT 3.1 outperforms other models in various benchmark tests, achieving high scores in both deep and non-deep thinking modes across multiple assessments [4][5] - In deep thinking mode, MindGPT 3.1 scored 0.8625 in AIME 2024, indicating strong performance relative to competitors [4] Learning Methodology - The ASPO method addresses the issue of data sampling precision, focusing on filtering low-quality learning signals to enhance model training [8][9] - Unlike GRPO, which operates at the output stage, ASPO manages the training pool at the input stage, ensuring that only samples that match the model's capability are used [8][9] Strategic Focus - Li Auto's leadership emphasizes that the primary focus is on enhancing model capabilities rather than artificially inflating benchmark scores, which they consider a waste of resources [5][6] - The company is committed to improving user experience by reducing reasoning time and enhancing the overall quality of responses from the model [5] Collaborative Initiatives - Li Auto has initiated a joint fund with local scientific committees to engage with academic professionals, aiming to gather the latest research insights without specific deliverable requirements [10]

Core Viewpoint - DeepSeek officially released DeepSeek-V3.1, marking a significant step towards the "Agent era" with enhanced capabilities in reasoning and task performance [1] Group 1: Product Upgrade - The upgrade includes a mixed reasoning architecture that supports both thinking and non-thinking modes in a single model [1] - DeepSeek-V3.1-Think can provide answers in a shorter time compared to its predecessor, DeepSeek-R1-0528 [1] - The new model shows significant improvements in tool usage and intelligent agent tasks through Post-Training optimization, resulting in stronger agent capabilities [1] Group 2: User Experience - The official app and web model have been synchronized to DeepSeek-V3.1, allowing users to switch freely between thinking and non-thinking modes via a "deep thinking" button [1]

Core Insights - DeepSeek has officially launched DeepSeek-V3.1, marking a significant step towards the era of intelligent agents with its hybrid reasoning model and 671 billion parameters, surpassing previous models like DeepSeek-R1 and Claude 4 Opus [1][12][18] Model Performance - DeepSeek-V3.1 demonstrates faster reasoning speeds compared to DeepSeek-R1-0528 and excels in multi-step tasks and tool usage, outperforming previous benchmarks [3][6] - In various benchmark tests, DeepSeek-V3.1 achieved scores of 66.0 in SWE-bench, 54.5 in SWE-bench Multilingual, and 31.3 in Terminal-Bench, significantly surpassing its predecessors [4][15] - The model scored 29.8 in the Humanity's Last Exam, showcasing its advanced reasoning capabilities [4][16] Training and Architecture - The model utilizes a hybrid reasoning mode, allowing it to switch between reasoning and non-reasoning modes seamlessly [6][12] - DeepSeek-V3.1-Base underwent extensive pre-training with 840 billion tokens, enhancing its contextual support [6][13] - The training process involved a two-stage long context expansion strategy, increasing the training dataset significantly [13] API and Accessibility - Starting September 5, a new API pricing structure will be implemented for DeepSeek [7] - Two versions of DeepSeek-V3.1, Base and standard, are available on Hugging Face, supporting a context length of 128k [6][14] Competitive Landscape - DeepSeek-V3.1 has been positioned as a strong competitor to OpenAI's models, particularly in reasoning efficiency and coding tasks, achieving notable scores in various coding benchmarks [12][20][23] - The model's performance in coding tests, such as Aider, reached 76.3%, outperforming Claude 4 Opus and Gemini 2.5 Pro [16][19]