Alibaba announced a multi-layer upgrade to its AI stack at the Alibaba Cloud Summit, covering cloud infrastructure, model services, AI chips and foundation models to support customers building and deploying AI agents.
New foundation model
The company introduced Qwen 3.7-Max, a large language model tailored for agentic coding, complex reasoning and long-horizon task execution. Alibaba said the model will be made available soon to developers and enterprises worldwide through its model service platform, Model Studio.
According to the company, Qwen 3.7-Max is capable of code generation and debugging, automating office workflows and managing complex multi-step procedures that require hundreds or thousands of actions. Alibaba stated that the model can autonomously execute tasks for up to 35 hours and handle more than 1,000 tool calls without degradation in performance. The model is optimized for a range of agent frameworks, including OpenClaw, Hermes Agent, Claude Code, Qwen Paw and Qoder.
Server and system-level hardware
To support agent inference and large-scale model training, Alibaba Cloud launched the Panjiu AL128 Supernode Server. The system integrates 128 AI accelerators within a single rack and delivers single-rack bandwidth at the petabyte-per-second scale, according to the company. Alibaba Cloud said the Panjiu AL128 is available on Model Studio for the China market.
Processor and networking silicon from T-Head
T-Head, Alibaba's semiconductor design unit, unveiled the Zhenwu M890, an AI training and inference processor that Alibaba says provides three times the performance of its predecessor, the Zhenwu 810E. The M890 offers 144 GB of GPU memory and 800 GB per second of inter-chip bandwidth, and supports a range of numerical precision formats from FP32 down to FP4.
T-Head also introduced the ICN Switch 1.0, a switching chip capable of delivering up to 25.6 Tbps of aggregate bandwidth while enabling full-bandwidth interconnection across 64 accelerators.
On deployment reach, Alibaba reported that T-Head has delivered more than 560,000 Zhenwu units to date, with over 400 external customers across 20 industries using the chips.
What this means
The announcements span software and hardware layers designed to help enterprises adopt agentic AI workflows. Alibaba positions the combined offering as a vertically integrated stack that includes foundation models, model hosting and high-bandwidth compute and networking hardware to support extended-duration agent operation and heavy tool invocation.