Recently, Alibaba Cloud officially announced the full launch of its Lingjun Zhenwu M890 super node instance, with the first batch already available for sale in the Ulanqab region. The release of this significant computing product marks a key breakthrough in the field of large-scale model training and inference in China.

According to the information, this super node instance is the first in China to successfully run a large model with over 2 trillion parameters. Currently, well-known large models such as Kimi K3 and Alibaba's latest flagship model Qwen3.8 Max (with a parameter scale of up to 2.4 trillion) have already been made available through this instance.

In terms of hardware specifications and technological innovation, the Lingjun Zhenwu M890 super node instance is equipped with the next-generation training and inference integrated AI chip from T-head, the Zhenwu M890. This chip supports multiple data precision levels, ranging from FP32 to FP4. Combined with the ICN Switch 1.0 interconnection chip, it achieves a high-speed interconnection of 800GB/s between 64 cards, with a memory pool capacity of 9TB. A single instance can easily support the expert parallel requirements of ultra-large-scale 2-trillion-parameter MoE models.

Thanks to the full-chain deep optimization of the underlying software and hardware, this super node natively supports low-precision computing with FP8/FP4. In training scenarios such as intelligent driving and embodied intelligence, the performance across multiple scenarios has increased by up to three times compared to the previous generation Zhenwu 810E. It can also achieve a performance improvement of up to 1.5 times in Agentic reasoning scenarios. Enterprise customers no longer need to build their own data centers; they can directly activate high-speed interconnected computing units on the cloud. A single instance can support the inference of MoE large models with up to 100 trillion parameters.

On the underlying intelligent computing support, the Lingjun Zhenwu M890 super node relies on the unified base of the Lingjun Intelligent Computing Platform and is equipped with the HPN 8.0 training and inference integrated network. Its single cluster can support up to 130,000 heterogeneous computing resources, and can be flexibly expanded to millions of cards, providing a solid computing foundation for the current explosive growth of large-scale AI workloads.