Alibaba Cloud's Zhenwu M890 Supernode Adapts Qwen3.8, Goes Live on Bailian for Inference

On July 23, Alibaba announced that its Zhenwu M890 supernode has successfully adapted Qwen3.8 and is now live on the Alibaba Cloud Bailian platform for model inference — the first supernode in China to run a large model with over 2 trillion parameters, as reported by QbitAI.

Qwen3.8 is Alibaba's latest flagship model with 2.4 trillion parameters. Running a model of this scale requires distributing parameters across dozens or even hundreds of high-speed interconnected GPU cards for high-throughput, low-latency collaborative computation. Conventional AI clusters, constrained by limited communication bandwidth, cannot meet these demands.

The Zhenwu M890 is Pingtouge's next-generation training-inference integrated AI chip, natively supporting precision levels from FP32 to FP4, covering everything from high-precision training to ultra-low-precision inference. The supernode connects 64 Zhenwu M890 chips through the ICN Switch 1.0 interconnect, achieving 800GB/s chip-to-chip bandwidth with 9TB of total memory — sufficient for expert parallelism in a 2-trillion-parameter MoE model.

On the software side, Alibaba implemented full-stack collaborative optimization from the chip layer to the cloud platform for Qwen models. This allows Qwen3.8 to run smoothly on the Zhenwu supernode while significantly improving inference efficiency and reducing costs. Test data shows up to 1.5x inference performance improvement in Agentic reasoning scenarios through operator optimization and hardware-software co-design.

The Zhenwu M890 supernode's adaptation of Qwen3.8 demonstrates that China's domestic AI chip ecosystem is maturing rapidly. Pingtouge's transition from early training-focused chips to commercial deployment in ultra-large-scale inference scenarios marks a critical step forward for domestic AI hardware in the premium market segment.

For Alibaba Cloud, connecting the Zhenwu supernode with the Bailian platform creates a full-stack loop from chip to model to cloud service. This means Alibaba Cloud not only provides compute power but has deeply integrated capabilities across the chip, interconnect, and platform layers, building a higher competitive moat.

With Qwen3.8's inference service now officially live on Bailian, enterprise users can directly call this domestic ultra-large-scale model, running on homegrown chip infrastructure. This provides a self-controllable option for the large-scale deployment of domestic AI applications.