Alibaba Unveils Agentic Cloud and Qwen 4 Roadmap

Alibaba used its annual Apsara Conference in Hangzhou to lay out a full-stack artificial intelligence roadmap, announcing a purpose-built agentic cloud, an enterprise platform for AI agents, next-generation Qwen models, custom AI chips and a line of AI-powered devices. The announcements were made under the conference theme "Intelligence Goes Beyond" and were detailed in a company announcement distributed on September 22.

The roadmap covers nearly every layer of the AI business, from the data centers that train models to the smartphones people carry. Chairman Joe Tsai framed the theme as "guiding AI from technological breakthroughs toward value creation." Chief executive Eddie Wu said machine thinking currently amounts to less than 3 percent of all human thinking, and that Alibaba Cloud is targeting more than 20 gigawatts of global data center capacity by 2032 to meet rising demand for AI compute.

A cloud built for agents

The centerpiece of the announcement is a shift from selling raw models to selling managed agent infrastructure. Alibaba introduced its agentic cloud, called Agent Native Cloud, along with AgentCore, an enterprise platform for building, running and managing AI agents across business systems such as customer service, coding and data analytics. An Agent Security Center bundles lifecycle management with security controls so companies can pilot agents while keeping governance centralized, according to coverage of the event.

Alongside the platform, the company described Agent Context, a system that gives agents real-time context and long-term memory. In knowledge-intensive work, Agent Context can cut token usage by up to 67 percent, which matters because token costs are one of the main expenses of running agents at scale. The pitch is aimed at chief information officers who want production-grade agents without stitching together their own infrastructure.

Larger models and custom silicon

On models, Alibaba confirmed that its next-generation Qwen 4 foundation model is in training, with future Qwen 4.5 and Qwen 5 releases expected to scale between five trillion and ten trillion parameters. The company also showed off what its current models can already do. Qwen 3.8-Max recently ran a month-long self-improvement loop across 33 cycles, raising its score on the Artificial Analysis Intelligence Index from 40 to 45, as reported by TECHz Magazine.

In a separate experiment, the model was given more than ten thousand electronic design automation calls to redesign a production-grade bus module. It reduced the module's area by 42 percent without any loss in performance, according to the same report. The company framed the experiments as demonstrations of autonomous capability.

The hardware side of the roadmap comes from T-Head, Alibaba's chip unit. It introduced the Zhenwu V900 AI processor, which launches in the first quarter of 2027 with triple the performance of the M890. The chip carries 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth with native FP8 and FP4 support, and it will power a new supernode server designed to scale to clusters of half a million accelerator cards. T-Head also laid out a CPU roadmap, with the Yitian 720 arriving the same year and the Yitian 750 following in 2029 with direct interconnections to Zhenwu AI chips.

From data centers to phones and laptops

The roadmap extends to devices. Alibaba introduced Qwen Intelligence, a full-stack agentic AI platform that combines mobile-optimized foundation models with an agent platform and ready-to-use AI agents. It is designed to move phones beyond assistants that answer one prompt at a time toward systems that plan and carry out multi-step tasks across apps, such as booking a trip and adjusting the itinerary if a flight is cancelled.

"Agentic smartphones are the primary gateway bridging personal AI and the physical world. Qwen Intelligence integrates mobile-optimized Qwen models with a custom-engineered harness, empowering phone makers to deliver seamless, personal intelligence to their users," said Steven Hoi, head of Qwen Mobile AI at Token Foundry and vice president of Alibaba Token Hub, as quoted in the announcement coverage. HONOR is the first manufacturer to adopt the platform, and its Magic9 series, launching September 28, will be among the first phones to ship with it, alongside the HONOR Robot Phone, as reported by Arabian Reseller.

Alibaba also unveiled Qwen Book, an agentic computer running Qwen Desktop OS, an AI-native operating system that folds Qwen models, applications and cloud services into one environment. The company described an "OS as Harness" approach in which the operating system exposes interface context and runtime controls so an agent can understand a task and execute it while staying governed, as reported by Technode. In a demo, a user gave voice instructions through a stylus while editing a presentation, and the system retrieved the file from the cloud, located the slide, invoked editing tools and prepared changes for review. New AI wearables, including smart glasses, AI earbuds and an upgraded portable note-taking device, round out the hardware announcements.

The multimodal lineup got updates too. New additions include Qwen-Audio-3.1-TTS-Next for cinematic soundscapes, Qwen-Image 3.1 for advanced image generation and the HappyOyster 2.0 Preview world model. The announcements span silicon, cloud infrastructure, models, agent platforms and the devices that reach end users. For the full list of announcements, see Alibaba's official roadmap announcement.