Author | Li Shuiqing, Editor | Xin Yuan
Reporting from Hangzhou on September 22, Zhidx reported that today, at the 2026 Apsara Conference, Alibaba Group CEO Eddie Wu announced a series of achievements by Alibaba in three major fields: AI models, AI chips, and AI Cloud, covering more than a dozen new products.
In terms of AI chips, Alibaba's T-Head launched the Zhenwu V900, which delivers three times the computing power of the Zhenwu M890. A single cluster can scale up to 500,000 cards and is expected to be officially ready in Q1 2027. Eddie Wu stated that this is currently the most powerful AI chip in China in terms of computing performance.
Also unveiled at the same time was the Panjiu Hyper-Node Server, which is equipped with Zhenwu V900, ICN Switch, Panjiu, and Zhenyue chips, and is also scheduled for release in Q1 2027. In terms of CPUs, the Yitian 720 and Yitian 730 are expected to be launched in Q3 2027. In terms of AI models, Alibaba released three new multimodal products in one go: the image generation model Qwen-Image-3.1, the music model Happy Shrimp 1.1, and the world model Happy Oyster 2.0 Preview.
The Qwen team also previewed the Qwen4 series, which includes Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B. Future versions such as Qwen4.5 and Qwen5 will scale up to 5-10T parameters. Liu Daheng, head of the Qwen large model, made his first public appearance in his new role.
In terms of AI Cloud, Alibaba Cloud CTO Jeff Li introduced the enterprise-grade Agent building and governance platform AgentCore, and the new-generation high-performance storage CPFS, along with a series of new products including Agent Sandbox, Tair KVCM, KV CacheStore, TPN, and Context Engine.
Alibaba announced that by 2032, the scale of global data centers operated by Alibaba Cloud is expected to exceed 20GW. In addition, it is worth mentioning that Fang Fei, President of the Honor terminal product line, announced at the conference that the Honor Magic9 series, to be released on September 28, will be the first official models equipped with Alibaba's AI smartphone solution Qwen Intelligence, with Honor Robot Phone also supported simultaneously.
Combined with previous reports on the cooperation between Qwen and Apple Intelligence, we believe that Alibaba's role is changing: it is no longer just a supplier of output models, but is also extending towards being a provider of terminal AI infrastructure. In terms of AI hardware, Alibaba also launched the Qwen AI Glasses N11 and N1 Pro, and Qwen AI ear-clip headphones, which are now open for pre-order.
Ping An Insurance (Group) Company of China Chief Technology Officer Wang Xiaohang also shared the cooperation results. Over the past year, Ping An's average daily Token usage increased from about 30 billion to over 300 billion, while computing power increased from 800 PFLOP to 1500 PFLOP during the same period, achieving 10 times the Token growth with less than 2 times the computing power growth.
01.Eddie Wu: Machine Thinking Will Surpass Human Thinking by 1000 Times Alibaba's Three Cornerstones Revealed
Eddie Wu noted that machines are becoming the main force for thinking, and intelligence is becoming a scaled product. Regarding the future of machine intelligence, he made two important judgments: 1. In the future, the total volume of machine thinking will reach more than 1000 times that of human thinking. Top-tier thinking will transform from a luxury good into a commodity available for scaled supply. 2. The representative products of the machine intelligence era have not yet emerged. Today's AI Coding only replaces existing human work. When machine intelligence explodes, it is hard to imagine how many new inventions will appear.
The era of machine intelligence also requires robust infrastructure. Eddie Wu stated that the three cornerstones of the machine intelligence era are AI models, AI chips, and AI Cloud, which are the prerequisites for the scaled supply of machine thinking.
Centered around the three cornerstones, Alibaba announced its latest progress and goals: 1. In terms of AI models, the Qwen team has made progress in RSI (Recursive Self-Improvement). Future models will scale up to 5-10T parameters, with the goal of completing more complex and longer-horizon tasks, moving towards ASI. 2. In terms of AI chips, T-Head released its most powerful domestic AI chip, Zhenwu V900, with computing power increased to 3 times that of Zhenwu M890, and a single cluster scalable to 500,000 cards; the annual shipment volume of T-Head AI chips is expected to increase significantly. Alibaba has independently developed multiple series of AI chips, including Panmai, Yitian, Zhenwu, ICN Switch, and Zhenyue.
3. In terms of AI Cloud, current customer AI demand is very strong, and medium- to long-term demand far exceeds supply capacity. Alibaba will go all out with all partners to invest in AI infrastructure construction, with the goal of exceeding 20GW in the scale of global data centers operated by Alibaba Cloud by 2032.
02.Alibaba Launches New Zhenwu AI Chip: 3x Performance, Ready in Q1 2027
Alibaba's T-Head launched the latest AI chip, Zhenwu V900, with computing power increased to 3 times that of Zhenwu M890, video memory reaching 216GB, inter-chip interconnection reaching 1200GB/s, and a single cluster scalable to 500,000 cards.
In his speech, Eddie Wu mentioned that Zhenwu V900 is currently the most powerful AI chip in China in terms of computing performance. Due to the maturity of T-Head's chip product line and its widespread application by customers, the annual shipment volume of T-Head AI chips is expected to increase significantly. Gao Hui, Vice President of T-Head Semiconductor, revealed that Zhenwu will adhere to an annual evolution pace. Zhenwu V900 is expected to be officially ready in Q1 2027. In Q3 2028, T-Head is expected to launch Zhenwu J900, with further breakthroughs in its self-developed parallel computing architecture and performance.
Based on Zhenwu, T-Head also announced the launch of the brand-new Panjiu Hyper-Node Server, equipped with T-Head's Zhenwu V900, ICN Switch, Panjiu, and Zhenyue chips, expected to be released in Q1 2027.
Based on T-Head's hyper-node, it supports native memory semantics, achieving unified memory addressing; it supports thousand-card full-bandwidth Scale-UP interconnection and hundred-TB-level unified video memory.
Gao Hui revealed that hyper-node instances based on Zhenwu M890 have been providing external services, marking the first hyper-node form factor computing power in China to successfully run a large model with over 2 trillion parameters. In terms of software, since the T-Head SAIL SDK was officially announced as open-source in July, its core components have taken the lead in completing open-source release today. The T-Head SAIL SDK supports minimal code modification, aligns ecosystem evolution with industry evolution, and natively adapts to all frameworks, allowing developers to avoid modification, waiting, and vendor lock-in. In addition to the Zhenwu AI chip series, T-Head also announced that the Yitian 720 and Yitian 730 CPUs are expected to be launched in Q3 2027, followed by the Yitian 750 equipped with the second-generation self-developed core.
03. Liu Daheng's Debut: Previewing Qwen4 Qwen3.8 Self-Evolution Iterates 33 Rounds in One Month Cost Reduced to 1/9
Liu Daheng, head of the Qwen LLM project at Token Foundry, Alibaba's ATH Business Group, made his first public appearance in his new role. Liu Daheng previewed the Qwen4 series, including four models: Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B. What is Qwen's North Star goal? Liu Daheng discussed two routes: first, pursuing the ceiling of intelligence through the Qwen-Max series; second, continuously lowering the usage threshold of models through series such as Qwen-Flash, Qwen-Plus, and Qwen-27B.
He noted that two months after the launch of Qwen3.8, real user growth reached 8.5 times, and Token usage increased by 12 times. The model is penetrating the real ecosystem, and its capability boundaries are rapidly expanding from general Q&A to more professional task domains. The Qwen team is accelerating the exploration of RSI. One month after the launch of the Qwen model self-evolution system, it has cumulatively completed 33 rounds of effective data iteration, driving significant improvements in the capabilities of the Qwen-3.8 model.
Currently, Qwen has been implemented in production-grade chip design and inference adaptation. In chip design, Qwen3.8-Max starts from an NoC module specification to complete RTL, verification, and physical implementation; in another independent experiment, Qwen3.8-Max deployed Qwen3.8-Flash-Next to T-Head's Zhenwu M890 and continues to optimize end-to-end inference performance.
The Qwen team is also focusing on innovating the model structure to achieve efficiency improvements. It is reported that the brand-new architecture upgrade of Qwen reduces the activated parameters of Qwen3.8-Flash to 1/3 compared to Qwen3.7-Plus, lowers training costs to 1/9, and reduces the price of one million input Tokens to 1/4.
In addition, the Qwen team is also broadening the boundaries of Agents through omni-modality and continuously breaking through the upper limit of intelligence via Scaling. Qwen4, based on the new-generation architecture, is already in training, and future versions such as Qwen4.5 and Qwen5 will scale up to 5-10T parameters.
Today, the global downloads of Qwen models exceed 3 billion, with over 460 models cumulatively open-sourced and more than 300,000 derivative models.
04.Alibaba Releases Three Multimodal Models in One Go Predicts Unified Model to Emerge Within Three Years
Zheng Bo, Vice President of Technology at Alibaba's ATH Business Group and Chief Scientist of Taobao Tmall Group, noted that omni-modal generation is accelerating Scaling, redefining people's experiences of "thinking" and "seeing".
Zheng Bo announced that Alibaba officially launched the Qwen-Image-3.1 model today. This model can meet complex generation needs in various scenarios, such as movie script rendering, academic knowledge rendering images, and detailed introductions of consumer products; its image editing capabilities have been upgraded, allowing the generation of multi-dimensional visual content such as sketches and surrounding environments from a single selfie.
At the same time, Alibaba launched the Happy Shrimp 1.1 music model. This model supports over a hundred music styles and more than a dozen mainstream languages, with comprehensive upgrades in capabilities such as melody, arrangement, vocal emotion, understanding, control, and multilingual support, making it more suitable for real content consumption scenarios.
In terms of world models, Alibaba launched Happy Oyster 2.0 Preview, with improvements in physical consistency, memory accuracy, and fast response capabilities compared to the previous generation model.
Zheng Bo predicted that within three years, a native omni-modal unified model will emerge, allowing experiences to no longer be limited by the boundaries of the model.
05.Agent Explosion Forces Cloud Base Upgrade Alibaba Cloud Launches AgentCore New-Generation CPFS and Other Products
Alibaba Cloud CTO Jeff Li systematically elaborated on the Agentic Cloud strategic layout. Alibaba Cloud takes Model, Harness, and Context as core construction scenarios, launching a series of new products including a brand-new cloud computing power base, the Agent building and governance platform AgentCore, Agent Sandbox, and the new-generation high-performance storage CPFS, enabling AI to truly enter production processes and move towards scaled value creation.
The explosive growth of Agents poses higher requirements for computing power and cloud. Centered around the four capability pillars of omni-domain cognition, autonomous operation and continuous evolution, trusted collaboration and governance, and real-time intelligent decision-making, Alibaba Cloud has reconstructed the AI Cloud technology stack system.
At the conference, Alibaba Cloud showcased its full-stack self-developed computing power base, covering multiple product lines such as computing, storage, and networking, to address the challenges brought by RSI. The Lingjun Zhenwu V900 hyper-node, which will be launched next year, will triple performance, with a single cluster scalable to 500,000 cards, evolving towards wide-area hyper-nodes. At the same time, Alibaba Cloud also launched the new-generation full-stack self-developed high-performance storage CPFS, providing up to hundred-TB/s-level throughput and hundred-million-level IOPS, increasing the scale of a single file system by 5 times to 100PiB, supporting large-scale model training and multimodal data processing.
In actual model training, it can reduce the average time for model startup by 50%, increase peak computing power utilization by 30%, and reduce AI storage costs by 69%, shifting GPUs from "waiting for data" to continuous and efficient computing. In terms of inference, Alibaba Cloud launched new products such as the large model KVCache management and scheduling system Tair KVCM, the storage acceleration engine KV CacheStore, and the brand-new intelligent computing center converged network TPN. Centered around the technological innovation of multi-party collaboration in "computing, networking, and storage", it reduces costs and increases efficiency for model inference.
In addition, the artificial intelligence platform PAI has added support for an asynchronous AgenticRL training framework, connecting the loop of trajectory sampling, reward accumulation, model training, and parameter updating. In practical applications supporting the post-training of Qwen models, it achieves completing one post-training of a SOTA model in 5 days. To enable enterprise-level customers to deeply collaborate with Harness faster, Alibaba Cloud launched the enterprise-grade Agent building and governance platform AgentCore, standardizing and hosting enterprise-grade Agent infrastructure, supporting long tasks, failure retries, breakpoint recovery, and asynchronous execution; it provides a secure execution environment, resource isolation, identity permissions, and full-link auditing, uniformly managing assets such as models, MCP Servers, and Skills, which can improve task completion rates to 99% and reduce TCO costs by 70%.
In addition, Alibaba Cloud has also prepared a series of Agent-related products, including Agent Sandbox, Agentic OS, Agentic Computer, etc.; and proposed the trinity engine of "storage-computing processing-context engineering"—Context Engine, transforming enterprise omni-domain data into real-time context, enabling Agents to further move from "knowing what" to "what should be done right now".
06.Honor's New Phone Debuts Qwen Intelligence Ping An Supports 10x Tokens with Less Than 2x Computing Power
Fang Fei, President of the product line at Honor Terminal Co., Ltd., attended the conference and announced that the new Honor Magic9 series, to be released on September 28, will be among the first official models equipped with Alibaba's AI smartphone solution Qwen Intelligence, with Honor Robot Phone also supported simultaneously. It is introduced that the two parties joined hands to build domain-specific models and solutions for next-generation AI smartphones based on Qwen Intelligence and Honor MagicOS. Based on this solution, the comprehensive task accuracy rate reaches 91.8%, the GUI operation speed is only 3.6s, it can achieve 100+ operation steps for complex long-horizon tasks, and the end-to-end service closed-loop rate exceeds 90%.
Wang Xiaohang, Chief Technology Officer of Ping An Insurance (Group) Company of China, interpreted Ping An's AI cooperation with Alibaba in fields such as finance and healthcare. Over the past less than a year, with the support of Alibaba Cloud's AI full-stack technology optimization, Ping An's average daily Token usage increased from about 30 billion to over 300 billion, while computing power increased from 800 PFLOP to 1500 PFLOP during the same period, supporting 10 times the Token growth with less than 2 times the computing power growth. Through full-stack optimization of models, inference, and computing power scheduling, Ping An's Token output per unit of computing power increased by 368%, and the cost per Token decreased by 76%.
07.Conclusion: Agent Explosion Forces Cloud Base Reconstruction
From the three cornerstones of models, chips, and cloud announced by Alibaba, we can see that AI competition is expanding from single-point model capabilities to the scaled supply of full-stack infrastructure. The Agent explosion is already forcing the reconstruction of the cloud base, and the implementation data in terminal and financial scenarios also proves that scaled value creation is not just a distant vision. When models begin to iterate themselves, chips evolve annually, and the cloud expands towards wide-area hyper-nodes, the supply curve of machine intelligence is just beginning to rise.