A：The 2026 Hangzhou Yunqi Conference grandly opened today on the banks of the Qiantang River. The opening ceremony was presided over by Du Xuliang, Deputy Secretary of the Hangzhou Municipal Party Committee and Mayor of Hangzhou, and was attended by numerous leaders, experts, scholars, and entrepreneurs, including Liu Fei, Member of the Standing Committee of the Zhejiang Provincial Party Committee and Secretary of the Hangzhou Municipal Party Committee, as well as Joe Tsai, Chairman of the Board of Alibaba Group, and Eddie Wu, Chief Executive Officer of Alibaba Group. The conference, themed “Yunxi Town Seals a Digital Pact,” aims to showcase successful AI applications across diverse industries and to establish a vital platform for technological exchange and industry collaboration.

A：From January to July this year, the revenue of Hangzhou’s core artificial intelligence industry reached RMB 316 billion, a year-on-year increase of 26.5%. Hangzhou has successfully established four national artificial intelligence application innovation bases, ranking first among all cities in China. Meanwhile, the performance of Hangchan Base’s open-source large language model remains world-leading; for example, Alibaba Cloud’s Jiuchao super-node server, which adopts an innovative architecture, was selected as one of the “Top Ten Treasures” at the 2026 World Artificial Intelligence Conference, and Weina Core’s three-dimensional computing-and-storage-in-memory chip architecture has also broken through multiple technological bottlenecks.

A：Looking ahead, Hangzhou will thoroughly implement General Secretary Xi Jinping’s important remarks on artificial intelligence, continue to carry out three major initiatives—open-source ecosystem development, breakthroughs in embodied intelligence, and demonstration of institutional innovation—as well as six foundational projects, including model advancement and computing power infrastructure, to accelerate the use of AI in empowering all sectors of the economy. In addition, Hangzhou will strengthen technological innovation, allocate stable funding to support basic research, build self-reliant and controllable AI software and hardware systems, and actively foster new forms of the smart economy. It will also deepen the application of AI in fields such as healthcare, education, science, and culture and tourism, striving to achieve a revenue scale of over 550 billion yuan for its core AI industries. In his speech, Joe Tsai, Chairman of the Board of Alibaba Group, stated that the Alibaba Cloud栖 Conference, as a platform for technological exchange, has witnessed continuous breakthroughs and large-scale applications of AI technology. Alibaba is firmly committed to building AI across its entire platform, creating a full-stack system that spans from chips to applications, enabling technological advancements to be more efficiently translated into industrial progress. In the new era of machine intelligence, Alibaba will remain committed to open collaboration, continuously build AI capabilities, innovate to serve the future, and contribute to the development of the artificial intelligence industry in Hangzhou and around the world.

A：Machine intelligence is not a substitute for human intelligence; it is an entirely different species. In the future, the significance of machine intelligence to human society will far exceed its current role as a substitute for mental labor. When the supply of machine intelligence becomes virtually limitless, its impact on society will be even more profound than that of the Industrial Revolution. On the supply side, human cognitive capabilities are limited, whereas once AI attains a professional research level, it will be able to provide intellectual services at scale, transforming top-tier thinking from a luxury good into a scalable commodity. On the demand side, as AI capabilities improve, goods and services are no longer constrained by population size but are directly proportional to the depth of machine intelligence and the scale of supply.

A：We have two main judgments. First, it is estimated that the total amount of computational power provided by machines will be more than 1,000 times the total cognitive capacity of humans. In the physical world, machines already handle nearly 99.9% of the work performed. In the future, the total volume of machine-based reasoning will expand at an even faster pace, eventually accounting for 99.9% of all cognitive activity in the world.

A：Yes, although AI applications are emerging one after another, early flagship products like the electric light can usher in an entirely new era. The truly iconic product of the era of machine intelligence has yet to emerge, much like the light bulb in its time—just one among Edison’s many inventions.

A：The era of machine intelligence requires robust infrastructure as its three cornerstones: AI models, AI chips, and AI cloud. Only when these infrastructures are sufficiently well-developed can they support the birth of various new AI products and inventions. AI models need to continuously iterate and evolve to handle more complex tasks; AI chips must enhance performance and expand supply to meet computational demands; AI cloud is responsible for producing and distributing tokens, building a highly efficient global network that enables devices to access machine intelligence anytime, anywhere. Alibaba will firmly invest in infrastructure development, including AI models, AI chips, and AI cloud, viewing it as a long-term strategic choice. Through continuous research and development and optimization, we aim to unlock the limitless potential of future AI and are committed to building a highly efficient, globally interconnected super network, making machine intelligence as readily accessible as electricity.

A：In the agency‑mode era, Qwen is focusing on four key areas: embedding the model deeply into real‑world industrial production, building a self‑refining loop through capability enhancement, scenario penetration, and feedback accumulation; in the coding domain, Qwen is accelerating its adoption into more complex programming tasks, driving autonomous development of sophisticated software systems and achieving a comprehensive leap in coding capabilities; as Qwen expands into additional real‑world industries, its ability to handle professional work has been validated, with results showing that it has entered the global top tier across multiple specialized fields; finally, we are exploring a closed loop for model training that enables self‑evolution, leveraging internal test logs and user feedback to automatically identify and optimize model shortcomings, thereby establishing a data‑driven self‑purification mechanism. Qwen-3.8 Flash achieves four key breakthroughs on its new architecture: it replaces the full-attention mechanism with a sparse-attention–mixed-linear-attention scheme, making it better suited for long-text tasks; it introduces a self-developed gate residual mechanism to enhance information flow and improve training stability; it incorporates n-gram embeddings to expand model capacity at a lower cost; and it employs a co-designed architecture and optimization approach to boost both training efficiency and model performance.

A：In the field of chip design, Qianwen has been integrated into a real-world industrial-grade EDA environment and is participating in DAMO Academy’s research on 3D innovative architectures. Using only a bus‑module specification as input, Qwen simultaneously iterates on front‑end verification and back‑end full‑flow optimization to reduce chip area and power consumption. After more than sixty hours of continuous operation, it executed tens of thousands of industrial‑grade EDA tool runs, ultimately achieving a 42% reduction in chip area while maintaining performance.

A：By automatically identifying the model’s current capability gaps from real-world internal testing logs and user feedback, we then automatically generate new training data and environment models to address these deficiencies. The in-house data generation pipeline first undergoes multiple rounds of validation and iteration on small models; after optimizing data quality, it feeds into the training of larger-scale models, establishing an automated, continuous, data-driven self-cleansing mechanism that enables the models to continuously identify their own shortcomings in real-world scenarios and drive iterative improvements in subsequent model generations.

A：In just 30 days, Qwen-3.8-Max achieved another performance breakthrough, climbing from 40 to 45 points on the Artificial Analysis AA Intelligence Index and reaching the top tier in China. Furthermore, the model’s capabilities have also been significantly enhanced across various internal and external complex agency task benchmarks.

A：For Qwen, being multimodal means not only understanding text, but also processing various modalities such as linguistic code, audio, images, and video, thereby truly becoming a general-purpose AI capable of comprehending real-world information. In the field of visual multimodality, Qwen has achieved remarkable results: it not only understands a single image but also helps users comprehend, explore, and create the world. Meanwhile, by enhancing audio–video reasoning and A- type task capabilities, and by reducing costs through intelligent density optimization, we are further advancing the development of general-purpose intelligence across all modalities.

A：Alibaba has progressively built a full-modal production capability that includes models such as Ten Million Image One, Happy House, Happy Strip, and Happy Oyston. Among these, Ten Million Image 3.1 is specifically designed for productivity-level image generation, enabling precise replication, design editing, and comprehensive layout—demonstrating powerful image generation and editing capabilities. In addition, there are the future-oriented long-video generation models One and Happy House, the professional-grade music creation model Happy String, and Happy World, which offers an interactive experience.

A：Happy Stream 1.1 features four key improvements: First, the melodies are more catchy, with stronger memorable hooks, enhanced musicality, and greater stability; second, the vocals are more lifelike, with diverse singing styles and finer emotional control; third, the model demonstrates deeper understanding, significantly improving its ability to interpret creative intent and complex instructions, resulting in better adherence to prompts; fourth, multilingual singing is more accurate, with clearer pronunciation and fewer typos or unclear pronunciations.

A：For professional creators and businesses, Happy Stream 1.1 is a sophisticated, highly controllable music production tool that helps them express their creative vision with greater precision. For the average user, just a single sentence—no music theory or instrumental skills required—can turn an idea into a complete song.

A：The Happy OS 2.0 preview version features three key improvements: First, the physics are more realistic, with scene objects behaving in accordance with real-world physical laws; second, enhanced 3D spatiotemporal memory ensures more stable scene recall; and third, AR diffusion and on-policy training significantly enhance the real-time interaction experience, delivering faster feedback and lower latency.

A：Video generation models have gone through the DIT validation phase, resolving quality issues. Starting in the first half of the year, they began evolving toward multimodal reference capabilities, enabling the understanding and generation of text, images, videos, and audio. Now, they are advancing into the agent-video stage, where models begin to comprehend creative goals and intentions, better grasp narrative and cinematography, automatically deconstruct plots, organize characters and settings, and ultimately generate high-quality video content.

A：Models moving toward agent video need to evolve in the directions of longer sequences, greater controllability, higher completeness, and increased intelligence. Longer means maintaining consistency while telling a complete story; controllable refers to the creator’s ability to precisely manage every creative element; complete signifies an understanding of the plot’s emotional tone and a cohesive narrative; and intelligent implies a deeper grasp of the creative intent, with a director’s mindset.

A：It is expected that within three years, a native, multimodal unified generative model will emerge—one model capable of generating images, videos, and music, while also understanding the world. This will eliminate modal boundaries, enabling seamless integration and large-scale deployment of experience production.

A：Pingtouge Semiconductor is reshaping computing power through three key efforts: first, shifting the focus from training-centric to inference-first; second, evolving workloads from single‑request processing to multi‑step orchestration, requiring systems to handle longer contexts and multi‑turn interactions; and third, expanding the competitive landscape from single‑chip metrics to system‑level collaboration, emphasizing full‑stack synergy and ecosystem development.

A：Alibaba Cloud’s Ling Jun super-node instance GP9A is the first in China to successfully run a large-scale model with over 2 trillion parameters, achieving full-stack collaboration across chips, servers, and cloud models to support large-scale cluster operations. It demonstrates China's AI chip industry's capability to transition from validating individual performance metrics to delivering large-scale systems.

A：After the sale SDK was open-sourced, developers can reuse their existing experience without modifying code, quickly adapt to updates from mainstream inference frameworks without waiting for custom adaptations, and are no longer tied to sale, supporting a variety of training and inference frameworks. Overall, the Sale SDK provides a software environment that offers full compatibility, rapid evolution, and seamless production-level migration.

A：In the agented era, the CPU has transitioned from a supporting role back to the critical path. When dealing with complex workloads—such as function invocations, code sandbox initialization, intermediate-state persistence, and the aggregation of multiple results—all these tasks are handled by the CPU, and each step directly impacts user experience and the critical execution path. Therefore, the CPU must feature high single-core performance, wide memory bandwidth, and a short interconnect between the GPU and CPU. Alibaba Cloud has planned a new generation of Yitian product line, including the Yitian 720 with 192 cores and 12-channel DDR, supporting throughput-intensive scenarios and delivering extreme cost efficiency; as well as the Yitian 730, which will serve as the first-generation self-developed high-performance core, supporting two-way simultaneous multithreading, with single-core performance improved over the previous generation, targeting latency-sensitive scenarios such as AI high-node applications, and providing deterministic single-core performance. In the future, the Yitian 750 will also be launched, featuring a second-generation self-developed core that further enhances performance and natively supports T-Head’s IN‑link interface.

A：Alibaba Cloud is committed to building intelligent production and load networks. By focusing on the three key elements for deploying enterprise-level intelligent agents—model capabilities, comprehensive, timely, and accurate contextual awareness, and a harnessing capability that enables continuous self‑iteration, self‑purification, and self‑regulation—Alibaba Cloud has established four core capability pillars: a full‑domain cognitive engine, support for autonomous operation and continuous evolution of intelligent agents, trusted collaboration and governance capabilities, and real‑time decision‑making capabilities. At the same time, to address the challenges of production intelligence—such as the growing size of model parameters, the rapid iteration of training frameworks and architectures, and the expansion of cluster scale—Alibaba Cloud has pioneered innovations in underlying hardware, including self-developed CPU and GPU chips, and has built high-performance networking and virtualization-accelerated CIPU technologies to support large-scale model training and the efficient deployment of intelligent agents.

A：Our key innovations are primarily reflected in three areas: First, in sandbox capabilities, we boast industry-leading cold‑start and hot‑start speeds, as well as rapid sandbox resumption from deep sleep states, and can instantiate a large number of sandboxes in an instant. Secondly, in response to the storage requirements of intelligent agents, we have designed a system that efficiently creates and manages small files, supporting read/write operations on up to 500 billion files, and has achieved isolation of files and computational resources among multiple agents. Finally, these capabilities rely on underlying foundational technologies such as computing orchestration, networking, storage, and monitoring, providing robust support for the construction and operation of intelligent agents.

A：We have built a unified agent authentication system to ensure the secure, trustworthy, and reliable operation of agents on Alibaba Cloud. This authentication system integrates over 300 cloud products on Alibaba Cloud, using a single, unified identity authentication mechanism to ensure that intelligent agents operating within enterprise workflows are protected from unauthorized access.

A：Agent Core is designed to provide reliable, cost-effective, and high-quality capabilities for building, running, and monitoring AI agents, significantly increasing the adoption rate of agent development while reducing the need for unplanned manual intervention. Through fine-grained, all‑dimensional monitoring and a real‑time feedback mechanism, operational costs of intelligent agents are reduced, enabling enterprises to manage the entire market landscape, coordinate multi‑dimensional workflows, and evaluate and optimize their intelligent agents.

A：To address challenges such as resource scheduling, long-context dialogue, multi-task sharing, and long-window task retrieval, we have built a well-structured technical architecture that includes an efficient memory-pooled caching layer, a remote KV cache storage layer, and an optimized network architecture. By leveraging these technical approaches, we have significantly improved cache hit rates, reduced costs, optimized I/O performance, and enhanced bandwidth utilization, thereby providing agents with robust inference capabilities.

A：We provide a dedicated runtime environment and end-to-end security throughout the entire lifecycle, including operating system optimization to support multi-workspace integration, cloud desktop environment deployment, and a one-stop intelligent agent security center. The Security Center can intercept and track potential security risks in real time, ensuring the agent’s safety and reliability during task execution.

A：Comprehensive, real-time, and accurate context is one of the key elements that enable an agent to deliver value. By integrating Alibaba Cloud’s robust database and big data platform capabilities, we have addressed key challenges in enterprise data management—ranging from structured to multimodal data, real-time processing, and cross-agent collaboration—providing enterprises with a comprehensive, timely, and accurate enterprise context engine. This empowers intelligent agents to better understand and adapt to user scenarios, thereby enhancing their value and operational effectiveness.

A：In the Damai app, users can easily set a ticket‑grabbing alarm to receive a five‑minute reminder before ticket sales begin. In addition, the system can provide users with a playlist of ten popular songs by Harlem Yu and automatically generate a practice schedule for each song. Starting at 8:00 PM tomorrow night, users will practice one song every two days until all songs are completed; these schedules will be added to the user’s calendar.

A：In a car environment, users can use voice commands to enable motion sickness prevention mode on their phone, adjust the screen brightness and volume, set an alarm for three hours later, and simultaneously send a DingTalk message to inform others that they have departed. Finally, the system will provide the user with navigation to the Oriental Pearl Tower; this entire sequence of operations can be completed seamlessly across multiple apps with just a single voice command.

A：Honor has developed the industry’s first system-level agent Honey architecture, which has been first implemented on the Honor ROG Phone and Magic series. These models also come pre-installed with Alibaba’s Qwen intelligence as part of the initial rollout. For example, Honor phones feature the “U Suggestion” function, which can sense user needs in real time and provide context-aware reminders, such as boarding pass information and access to various QR codes. At the same time, innovative interaction modes on Magic OS 11, such as eye-tracking, also demonstrate Honor’s exploration in terminal sensing capabilities and innovative interaction methods.

A：Honor is developing core sensing capabilities for its devices, envisioning a future where sensing technologies are integrated into wearable devices such as smart glasses, smartwatches, and earbuds, transforming them into peripheral, collaborative sensing tools for users. In addition, Honor has extended its AI capabilities to the physical world, such as its self-developed robot with embodied intelligent interaction, and by optimizing user experiences in scenarios like sports and health, education, and travel through AI, it is committed to building a human-centered, all-scenario ecosystem.

A：As an AI practitioner in the industry, Ping An Group has extensively applied AI technologies in customer service, financial services, and healthcare. Through AI-powered quick services, AI medical services, and the use of AI in financial advisory, product development, and customer service, it has achieved significant improvements in efficiency, cost reductions, and the creation of greater business growth opportunities. Meanwhile, Ping An and Alibaba are collaborating deeply on full-stack AI technologies. By leveraging cutting-edge AI and full-stack optimization techniques, they have significantly shortened the cycle from technological innovation to value realization, enabling the widespread adoption of model agents such as the Qwen series and Tongyi series models within Ping An, thereby driving the large-scale development of industrial AI.