A：This year’s conference has focused on a phenomenon: every year around the Spring Festival, a tech trend of unprecedented scale emerges. For example, in recent years the spotlight has shifted to ChatGPT, Sora, DeepMind, and others, while this year the Lobster Model has gained widespread attention thanks to its practical applications and agent‑based capabilities. This marks a transition in the agent‑intelligence space, moving from model‑centric reliance to an application‑driven paradigm.

A：The emergence of AI agents signifies that AI development has shifted from the model stage to the application stage; it will permeate all industries at an unprecedented pace, with the focus of competition shifting from intelligence to execution. Users place greater emphasis on whether an agent can accomplish practical tasks rather than its intrinsic level of intelligence. The agent can not only automatically write code and independently invoke tools, but also build new tools, which has led people to recognize that AI is not merely a chat tool, but is evolving into a digital employee and an agent. They can autonomously perform complex tasks and continuously enhance their capabilities through self-evolution.

A：Baidu has built a full-stack capability for the Nebula Agent, encompassing chips, models, intelligent agents, and optimization tools and capabilities. Among them, the development of large models follows an application-driven principle; for example, the ERNIE 5.1 model has performed exceptionally well in multiple evaluations and supports seamless integration of various AI capabilities, such as search, coding, and in-depth research, forming a unified portal called do mate to achieve efficient collaborative work.

A：It is necessary to stay focused on the core value proposition of intelligent agents, filter out noise, build consensus, and accelerate action. AI agents will transform user interfaces, trust dynamics, and organizational structures. Enterprises must adapt to these changes by adopting flatter management hierarchies and enhancing the efficiency of information flow, while also tracking daily active agents (DAA) as a key metric for gauging platform vitality.

A：Do Mate is capable of handling and executing the tasks of these three positions in a single engagement. First, it handles customer service tasks by autonomously accessing the inbox to process customer complaint emails, crafting emotionally resonant responses, and proposing compensation measures. Secondly, in sales‑related tasks, do mate can cleanse and standardize data from heterogeneous source spreadsheets, analyze sales performance, identify anomalous fluctuations, and generate in‑depth industry‑level sales reports. Ultimately, it provides recommendations on product assortment and promotional priorities for both e‑commerce channels and company‑operated stores. Finally, for marketing campaigns, do mate generates model-wearing renderings and high‑quality HTML‑formatted promotional articles to boost the launch of new products, while leveraging flash‑sale pages to drive sales.

A：Key features that enable Do Mate to tackle complex tasks include: a cutting-edge technical framework that excels at constraint enforcement, guidance, validation, and feedback; a robust skill ecosystem and comprehensive security mechanisms; and the ability to evolve continuously, becoming smarter with use. In addition, do mate integrates Baidu’s high-quality AI capabilities accumulated over many years, such as AI search, Instant Match, and Famo Encyclopedia, and can ensure that the model stays on course and maintains consistent, undistorted actions when executing complex tasks lasting for hours or even days.

A：Miaoda App is the first professional‑grade app‑building platform. Users can create web pages, mini‑programs, and mobile apps through natural‑language dialogue, with the entire development process requiring just one person, as 90% of the code is automatically generated by Miaoda’s AI engine. Users can move their mouths at any time and place, and their ideas can fall to the ground immediately. The MiaoDa app also supports multi‑user collaboration, end‑to‑end closed‑loop development, and enterprise‑grade stability, helping users rapidly build high‑quality applications and already enabling numerous developers to unlock business value.

A：In the AI era, designers’ work has become more diverse; for example, Ziroom leverages AI tools like MiaoDan to build home‑interior‑design applications on demand without writing any code. By integrating with the home‑decor module, users can simply upload a base image and select a design style to quickly generate renderings, turning their vision into reality—what you imagine is what you see, and what you see is what you get. As AI becomes more deeply integrated, frontline employees in enterprises can directly leverage AI capabilities to turn their business ideas into tangible growth and productivity gains.

A：Code proficiency is a foundational technical capability, playing a decisive role in advancing base model capabilities (such as natural language processing, multimodal generation, and creative writing) and in numerous AI-native applications. However, as AI advances—especially with the emergence of code agents like MiaoDa—people who don’t know how to code can also build applications, leading to the devaluation of code itself; some in the industry even believe that code will become free. This has disrupted the software industry, making one-time‑use or disposable software feasible, as production costs are extremely low, rendering even single‑use requirements worthwhile to develop.

A：In the past, the software industry was characterized by two significant barriers—high entry thresholds and high costs—leaving a vast array of long-tail and personalized demands unmet. But now, code agents like MiaoDa are breaking down these two barriers, lowering the entry threshold for applications, reducing development costs, and making one-off software a viable option. As a result, the software industry is being redefined, and its market size could potentially expand by a factor of ten—presenting a tremendous opportunity.

A：Digital human agents are the universal interaction interface of the AI era, endowed with expressive capabilities such as speech, facial expressions, and gestures; they can earn people’s trust, provide real-time feedback, and deliver emotional value. As digital human agents become increasingly sophisticated, their application scenarios continue to expand, extending from live‑commerce to areas such as advertising and marketing, content creation, and public services, ushering in a new phase of large‑scale deployment.

A：Baidu Yijing can generate high-quality video content lasting several minutes or even hours, while offering robust interactivity to meet users’ creative needs across a variety of scenarios, such as livestream‑driven e‑commerce, advertising and marketing, and content creation. It can generate any video centered on a human‑character subject with a single click, significantly boosting individual creators’ productivity and helping businesses seamlessly enter the global market by producing high‑quality, commercially valuable content.

A：By integrating interdisciplinary knowledge and model‑building capabilities, Baidu Famo can deeply understand and decipher the underlying principles governing product formulations and process parameter settings, thereby addressing the limitations of traditional approaches that rely on seasoned process experts while lacking systematic knowledge transfer. It can leverage scientific computing models to provide chemical enterprises with optimized process solutions and, in collaboration with the Asia Bay National Laboratory, has refined rice breeding and drought‑stress modeling, significantly enhancing detection efficiency.

A：During the May Day holiday, on its live production system, Qingdao Port adopted Baidu’s Famo technology and received highly positive feedback from operations staff. The efficiency of the ATOS system improved by 10.21%, with projections indicating it can handle nearly one million additional TEUs annually, helping Chinese products reach global markets. Moreover, strategy optimization not only improves individual production metrics but also supports multi‑directional, multi‑scenario simulations, serving as an intelligent assistant that promptly adjusts plans to aid business experts in decision‑making and enabling the reusable deployment of AI assets.

A：As a world-leading automated container terminal, Qingdao Port’s intelligent control system, A‑TOS, faces the challenge of performing extreme‑case simulations in three-dimensional space when determining vessel stowage plans. By collaborating with Baidu’s strategic planning team, starting from data reconstruction, focusing on the positional details of each container, establishing real-world constraints, and screening candidate solutions, we leveraged AI for holistic decision-making and conducted 248 rounds of self-iterative optimization, successfully reducing the number of toppled containers, increasing the biphasic coupling rate, and improving container‑outage balance, thereby enhancing the port’s overall input‑output ratio.

A：The Punish-Cat agent is capable of self-evolution and autonomous decision-making; it can self‑verify and optimize task outcomes, establishing a near‑closed loop that requires minimal human intervention and continuously seeks better solutions. At the same time, the requirements for AI’s verifiability and closed-loop operation also drive process reengineering and the codification of experience into verifiable context, enabling decision-making agents to move beyond single-use cases and evolve into enterprise‑wide intelligence capable of understanding and implementing industry logic while continuously seeking optimal solutions.

A：Enterprises must undergo self-evolution by delegating more authority and reducing micromanagement, accelerating alignment and flattening hierarchies, increasing talent density while moving away from labor‑intensive strategies, and integrating more tasks while minimizing rigid division of labor. In the AI era, corporate management must shift its paradigm, embrace innovation and creativity, and establish a systematic transformation geared toward the age of intelligent agents, in order to adapt to the disruptive impact of AI on how businesses operate.

A：The AI store manager can provide in-store guidance, review sales data, and, based on inventory levels, offer recommendations for product transfers and promotional strategies, while also issuing alerts for overdue orders. It comes with a suite of marketing tools to help businesses manage their operations.

A：Yes, the AI store manager’s capabilities can continuously expand and evolve. For example, once logistics data is integrated, it can calculate shipping costs and delivery lead times, helping merchants optimize their operations and management. We have made this capability available to our customers and partners. A local partner has leveraged Baidu Smart Cloud’s Hold‑On technology and related intelligent agents to build a solution for Yiwu merchants.

A：Baidu leverages its full-stack AI capabilities to participate in the intelligent transformation of the automotive industry, such as supporting Geely in building its first proprietary cloud and assisting Horizon in constructing an ultra-large-scale cluster. Baidu collaborates with leading companies across sectors—including automakers, battery and chip manufacturers, and autonomous‑vehicle providers—to co‑develop, train, simulate, test, and mass‑produce intelligent driver‑assistance solutions. Last year alone, it successfully supported the delivery of over 20 million new vehicles equipped with Level 2 advanced driver‑assistance systems.

A：Changan Automobile has leveraged the intelligent computing center it co-built with Baidu to establish core competitive advantages spanning from the smart cockpit to autonomous driving. At present, Changan Automobile has leveraged Baidu’s platform to significantly optimize the performance of its computing center, enabling efficient training and iteration of autonomous-driving models, and has become one of the companies approved by the Ministry of Industry and Information Technology to produce L3‑level autonomous vehicles.

A：Horizon’s strategic upgrade includes the launch of its new product line, Xingkong, an integrated chip designed for full‑vehicle computing. It adopts a central computing architecture, seamlessly integrating intelligent driving and smart cockpit processing, and supports a revolutionary human–machine interaction experience—such as the introduction of an in‑vehicle agent OS. Horizon has established a deep strategic partnership with Baidu AI Cloud, jointly building a full‑process closed loop based on big data, creating an ultra‑large‑scale heterogeneous computing cluster, supporting the distributed collaborative work of thousands of engineers, and raising industry benchmarks in the autonomous driving field in terms of scale, quality, and iteration speed. At present, Horizon Robotics has empowered more than 40 automakers worldwide, bringing hundreds of vehicle models to market, and is collaborating with companies such as the Volkswagen Group and Toyota.

A：Horizon has launched a high-performance autonomous driving chip with 560 TOPS of computing power, advancing intelligent driver-assistance technologies worldwide. Vehicles equipped with Horizon’s HSD have won favor among consumers thanks to their exceptional user experience, securing the top spot in China’s market with a share exceeding 30%.

A：AI is becoming a new super entry point by integrating with more hardware devices, such as smartphones and AI glasses. Taking Honor as an example, Baidu Intelligent Cloud provides it with end‑cloud integrated solutions and large‑model capabilities such as the ERNIE model, thereby enhancing user experience. For example, Honor UU has amassed over 100 million users, offering travel guides, attraction commentary, and personalized tour‑guide services, while also recommending local cuisine and providing navigation through Baidu Maps. At present, all of the world’s top ten smartphone manufacturers are leveraging Baidu Intelligent Cloud’s agent‑info capabilities. As open‑cloud general‑purpose agents gain traction, phone makers are beginning to expose OS‑level skills; when paired with agent‑building capabilities, this enables them to deliver a natively integrated super assistant for their devices.

A：When a user requests travel from Guangzhou to Chengdu during the May Day holiday, the AI assistant will invoke its built-in schedule management skill to check for time conflicts and, based on the user’s preferences, select suitable flights and hotels. After the user confirms the flight and hotel, the AI assistant will generate a final itinerary based on their choices, then use calendar and reminder skills to add it to the calendar and set a departure alarm, thereby achieving personalized trip planning.

A：Baidu Intelligent Cloud not only serves more than 1,000 AI hardware companies, providing services such as building core computing power models to enhance product experience, but also helps traditional hardware evolve into super-intelligent entities. Furthermore, in response to the challenges facing the embodied intelligence industry, Baidu Intelligent Cloud draws on its experience in the autonomous driving field, supporting large-scale data collection and annotation, accelerating model training and improving efficiency, reducing inference latency, and providing end-to-end capabilities that encompass computing platforms, toolchains, data services, voice interaction security, as well as training, evaluation, and feedback loop systems, thereby helping enterprises build competitive advantages in the era of physical AI.

A：Mr. Gao, CEO of Xinghai Map, shared his company’s experience in the entrepreneurial journey of Jusheng Intelligence, emphasizing that it is a well-rounded, integrated hardware‑software venture with no weak links, encompassing multiple stages such as motors, models, data, and toolchains. Jusheng Intelligence differs from large language models and autonomous driving in that its value chain is longer and every link is indispensable, necessitating an integrated hardware‑software solution. Xinghai Map adheres to an integrated hardware-and-intelligence strategy, placing particular emphasis on operational excellence as its core competency, and has independently developed the RE platform to serve developers worldwide. The Starry Sky Map focuses on two co‑evolving, symbiotic technical models—WAM and VLA—aiming to endow robots with general intelligence capable of understanding the world and manipulating all objects. They propose that pre-training should enable robots to learn the fundamental principles governing their own bodies and their interactions with the physical world, while fine-tuning serves as the adaptation process tailored to specific tasks. The world’s fastest world model, fast WEM, released in March of this year, has reduced one-step latency to 190 milliseconds, thereby enhancing deployment efficiency. The foundational model training for the Star Map is conducted on Baidu Intelligent Cloud, leveraging its efficient AI infrastructure to enhance development efficiency, with plans to further optimize model performance. Meanwhile, Xinghai Map has open-sourced its foundational models G0 and G0.5 and launched the Jusheng Intelligence one-stop development platform EDP, supporting end-to-end workflows that encompass data, model, device, and personnel management.

A：Financial operations place high demands on the professionalism, stability, and timeliness of AI services; particularly in an environment emphasizing self-reliance and controllability, it is essential to ensure that chip computing power, models, and applications have undergone rigorous testing. For example, China Merchants Bank has deployed the Kunlun Core P800 at scale and has already launched hundreds of AI applications covering core scenarios such as risk control, marketing, R&D, and office operations. SPDB has also leveraged AI to enhance its capabilities in refined operations and risk management. Baidu Intelligent Cloud has established deep partnerships with several financial institutions, including China Merchants Bank and Shanghai Pudong Development Bank, to jointly build a domestically developed heterogeneous computing cluster. The collaboration has enabled the evaluation of open-source models and the adaptation of mainstream large-scale models, resulting in the launch of specialized financial‑industry‑specific models tailored for applications such as financial‑statement‑form recognition and other vertical research areas. For example, in applications at Shanghai Pudong Development Bank, AI technology has reduced the time required for financial analysis from hours or days to just minutes, significantly boosting efficiency in risk management, marketing, and other areas.

A：Baidu Intelligent Cloud has collaborated with State Grid for ten years, advancing from intelligent customer service to the development of a trillion-parameter Guangming Power large model, successfully integrating AI technology into the grid’s core operations. For example, in the field of equipment inspection, combining small and large models has significantly improved recognition accuracy and reduced inspection time from 2.5 hours to 45 minutes, covering hundreds of substations nationwide and ensuring reliable power supply for 1.1 billion people.

A：Baidu Intelligent Cloud recognizes that customer demand for agent‑based applications has evolved, shifting from a focus on business elasticity and cost reduction with efficiency gains to an increasing need for highly active, high‑value, and scalable agent‑driven solutions. Therefore, Baidu Intelligent Cloud is focusing on building intelligent agents, comprehensively upgrading its product technologies, and developing flexible AI capabilities to address customers’ real-world business challenges.

A：Baidu Intelligent Cloud has comprehensively upgraded its capabilities in the AI info and agent info sections. AI info is committed to delivering the highest performance-per-watt and most cost-effective AI computing power, optimizing model performance; agent info, through token factory, enhances the speed and efficiency of agents in handling ultra-long-context tasks, and achieves more efficient agent operation via harness engineering. Meanwhile, by establishing collaboratively optimized infrastructure, agents can better adapt to complex task requirements and enhance their modeling and engineering capabilities.

A：Baidu Intelligent Cloud has achieved significant breakthroughs in underlying computing power, including its XunTui service optimized for intelligent agents, which delivers industry-leading memory reuse and heterogeneous scheduling technologies, as well as a multimodal training framework and reinforcement learning acceleration techniques, substantially boosting model training and inference efficiency. In addition, Baidu Intelligent Cloud has also launched the Kunlun Core P800 series, which deploys large-scale clusters on a large scale, and has enabled efficient and stable training and reasoning. It has also created high-performance and highly available AI computing infrastructure through optimized network architecture and data center design to promote the ecological development and technological innovation of intelligent bodies.