AI Affairs, home

Thursday 5 November 2026

Technology · Brief

Alibaba Cloud says its M890 SuperPoD supports Kimi K3.1 Pro from launch

The company says it has validated inference on vLLM and SGLang, while invite testing has opened for a Zhongwei configuration with a new networking card.

Alibaba Cloud entrepreneurship and innovation base sign outside a glass-fronted building
Photo: David290, CC BY-SA 4.0, via Wikimedia Commons (cropped)

On 5 November, Pandaily reported that Alibaba Cloud says its Lingjun Zhenwu M890 SuperPoD can run Moonshot AI’s Kimi K3.1 Pro from the release of its weights. The company says it has validated the model on the vLLM and SGLang inference frameworks.

Kimi K3.1 Pro has 2.8 trillion parameters, a sparse mixture-of-experts design and a context window of about one million tokens. The design draws on selected parts of the model for a request rather than activating every parameter. Alibaba Cloud, its chip unit T-Head and the Kimi team say they adapted the model using T-Head’s SAIL software stack.

Putting a newly released model to work could begin sooner when the infrastructure can run its weights on launch day, rather than requiring a separate adaptation period.

Alibaba Cloud has opened invite testing for a SuperPoD region in Zhongwei, Ningxia. Its configuration adds T-Head’s Panmai 920 networking card, which Alibaba says can reach 400Gbps. The company says the card’s built-in PCIe switch could cut system cost by about 30%.

Alibaba Cloud says over 100 major organisations use its M890 instances for AI model work, self-driving technology and intelligence embodied in physical systems. It expects its next-generation Zhenwu V900 SuperPoD instance to deliver about three times the performance of the current generation.

Sources

Topics: Chips, Enterprise adoption, Foundation models, Inference