Alibaba Cloud runs a huge AI model with 2.4 trillion parameters

Alibaba Cloud says its new Lingjun Zhenwu M890 super node can now run the Qwen3.8 model, offering AI inference services on its Bailian platform. This is the first super node in China to handle an AI model with over 2 trillion parameters.

Qwen3.8 is Alibaba's latest flagship model with 2.4 trillion parameters. The M890 super node uses ICN Switch 1.0 chips to connect 64 GPUs at high speed, giving it 9TB of memory. A single instance can handle the expert parallelism needs of a trillion-parameter MoE model.

Tests show that with hardware and software optimizations, the super node can boost inference performance by up to 1.5 times in agentic reasoning tasks while cutting costs.