Alibaba Unveils Full-Stack AI Strategy & Qwen Models

Joe Tsai, Chairman of Alibaba Group, framed the event around the theme Intelligence Goes Beyond, arguing that AI has expanded technological capabilities with vast potential for real-world deployment across thousands of industries.

Qwen 4 Enters Training

Alibaba confirmed its next-generation Qwen 4 model is now in training, with plans to scale Qwen 4.5 and Qwen 5 to between 5 trillion and 10 trillion parameters. The company also reported meaningful progress in recursive self-improvement, where models iterate on their own performance without human intervention.

Qwen3.8-Max completed 33 automated improvement cycles over a month, raising its Artificial Analysis score from 40 to 45. In a more unusual test, the model spent over 60 hours self-improving a chip design, making more than 10,000 electronic design automation tool calls. The resulting production-grade chip bus modules reportedly reduced chip area by 42% without sacrificing performance.

Multimodal Models Get an Upgrade

Alibaba’s multimodal lineup also expanded. New releases include:

• Qwen3.8-LiveTranslate, built for simultaneous interpretation
• Qwen-Audio-3.1-TTS-Next, which generates cinematic soundscapes combining dialogue and ambient sound from text
• Qwen-Image 3.1, aimed at creative design and e-commerce applications
• Updates across speech recognition, text-to-speech, and real-time interaction models

The company also launched Qwen Intelligence, a full-stack agentic platform for smartphones designed to handle complex tasks across multiple apps.

New Chips From T-Head

Alibaba’s chip design unit, T-Head, unveiled the Zhenwu V900 AI processor, built for both training and inference workloads. The chip claims three times the performance of its predecessor, the Zhenwu M890, with 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth. It supports FP8 and FP4 data precisions, with mass production and commercial release scheduled for the first quarter of 2027.

T-Head also introduced an upgraded supernode server integrating the Zhenwu V900, capable of supporting clusters of up to 500,000 cards. The company additionally outlined a roadmap for its next-generation Yitian 720 and Yitian 730 CPUs, both set for a 2027 launch.

Rethinking Cloud Around Agents

Alibaba Cloud’s agentic strategy centers on three layers: model, harness, and context. AI Native Cloud handles large-scale model training and inference, backed by updates to Alibaba Cloud’s Platform for AI, a new Cloud Parallel File Storage system, and HPN 8.0 Pro networking architecture.

Agent Native Cloud manages enterprise AI agent deployment, launching alongside AgentCore for building and managing agents and an Agent Security Center for compliance. The third layer, Context Engine, connects enterprise data to AI agents through a feature called Agent Context, which Alibaba claims can cut token usage by up to 67% in knowledge-intensive applications.

OpenLake, Alibaba’s data lakehouse product, was also upgraded into a unified multimodal system. The company claims the platform reduces total costs by 38% and query response times by 40% compared to traditional architectures.

Betting on Exponential Demand

Eddie Wu, CEO of Alibaba Group, pointed to what he called a significant growth runway for machine thinking, claiming the total volume of machine thinking currently sits at less than 3% of human thinking. Wu projects Alibaba Cloud’s global data center capacity will exceed 20GW by 2032, a buildout intended to meet what the company expects to be exponentially rising AI demand.

Hashlytics Take

The 3% figure is doing a lot of work here, and it is worth being clear about what it actually is. There is no established, verifiable way to measure “machine thinking” as a share of “human thinking,” so the number functions less as data and more as a framing device to justify a 20GW buildout eight years out. That does not make the infrastructure investment wrong. Cloud capacity planning genuinely does need to happen years in advance. But the parameter race toward 10 trillion for Qwen 5 and the self-improvement chip design results deserve more scrutiny than a press event allows, particularly since “42% smaller chip area” and “no performance compromise” are claims made by the company that built the model making the claim.