- The new Zhenwu M890 chip will handle both training and inferencing to meet the computing demands of the agentic era.
- Alibaba is advancing a comprehensive upgrade of its "chip-cloud-model-inference" stack and plans to update its AI chips annually.
Alibaba added a new processor to its expanding artificial intelligence (AI) technology stack, as the company endeavors to cover every aspect of AI development.
Its chipmaking unit T-Head unveiled the Zhenwu M890 AI accelerator at the Alibaba Cloud summit. Equipped with 144 GB of memory, the new hardware will handle both model training and inferencing jobs.
The chip is particularly suitable for handling high-concurrency agentic tasks. Its performance is three times that of its predecessor, with an interconnect bandwidth of 800 GB/s between chips, according to an announcement on Wednesday.
Alibaba Cloud also released the Panjiu AL128 supernode server based on the new chip at the summit. The server integrates 128 AI accelerators through a in-house developed interconnect chip.
Alibaba plans to upgrade its Zhenwu chips once a year going forward. This accelerates its update pace to roughly the same level as industry leader Nvidia.
The company plans to roll out two new generations of chips with stronger computing power, the Zhenwu V900 and Zhenwu J900, over the next two years to meet the AI computing demands of the agentic era.
T-Head has delivered a cumulative 560,000 units of its Zhenwu AI chip series, and these processors have been widely deployed by more than 400 external customers.
Major clients include large enterprises such as China Telecom, China FAW Group, and Shanghai Pudong Development Bank.
Beyond the underlying hardware, Alibaba also released its new-generation flagship model, Qwen 3.7-Max. The model is deeply optimized for complex tasks such as coding and multi-agent collaboration.
Large models are shifting from aligning with human preferences to aligning with task objectives, said Jingren Zhou, head of Alibaba's Tongyi large model division.
Furthermore, Alibaba Cloud launched a new AI product website, "Qianwen Cloud," which provides API access to more than 150 mainstream AI models.
The calling prices for some models on the platform are highly competitive, with one million input tokens costing as little as 1 yuan ($0.15).
To adapt to the cloud computing needs of the agentic era, Alibaba Cloud is comprehensively revamping its cloud products, a move aimed at providing developers with a more elastic and efficient operating environment.
($1 = 6.7921 yuan)
