August 12th news: Alibaba Cloud has officially launched the Lingjun Zhenwu M890 super node instance, which is now available for sale in the Wulanchabu region. Enterprise customers can now activate high-speed inter-connected computing units with 64 cards in the cloud without building their own data centers. This instance can support MoE large models with up to ten trillion parameters for inference.
According to the introduction, the Lingjun Zhenwu M890 is the first super node form of computing power in China that has successfully run a large model with more than 2 trillion parameters. Kimi K3 and Qwen3.8Max have already been provided through this instance. This instance supports FP8 and FP4 low-precision computing, and by using the ICN Switch1.0 chip, it has increased the scale of Scale-up interconnection from 16 cards to 64 cards, achieving an inter-card bandwidth of 800 GB/s. In training scenarios such as intelligent driving and embodied intelligence, the training performance is three times that of the previous generation Zhenwu 810E.
As a foundational computing platform, the Alibaba Cloud Lingjun Intelligent Computing Platform relies on the HPN8.0 training and inference integrated network, supporting up to 130,000 heterogeneous computing cards in a single cluster, and can be further expanded to millions of cards, providing large-scale computing power support for large model training and inference.
The official launch of M890 marks that super node computing power is extending further from self-built infrastructure to public cloud services. As trillion-parameter and even ten-trillion-parameter MoE models enter the stage of large-scale deployment, high-speed inter-card connectivity, low-precision computing, and large-scale cluster scheduling capabilities are becoming important indicators in the competition of AI infrastructure.
