Alibaba Cloud has completed day-zero adaptation for Moonshot AI's Kimi K3.1 Pro weights on its Lingjun Zhenwu M890 SuperPoD. According to IT Home, which cited the cloud provider on November 5, the company finished full-stack validation on the SGLang and vLLM inference frameworks.
The Kimi K3.1 Pro is described as a 2.8-trillion-parameter sparse mixture-of-experts model. It features a context window of about one million tokens, multi-token prediction, and hybrid Kimi Delta Attention and DSA sparse attention. These characteristics require high-speed scale-up interconnects and significant memory capacity.
Alibaba Cloud, T-Head, and the Kimi team collaborated to adapt the model on the T-Head SAIL software stack. This work included FlashMLA, DeepGEMM MoE operators, MegaMoE, and MoonEP. These developments allow customers to deploy inference on bare-metal SuperPoD instances or use the REIN intelligent engine service. The model is also accessible via the Qwen AI platform API.
Independent notes from the vLLM and SGLang projects describe the Zhenwu PPU support for Kimi K3.1 Pro on M890 and 810E configurations. These projects identified the release as the first time Zhenwu provided day-zero backing for a mainstream open-weight large model as an in-tree hardware backend.
Alibaba Cloud documentation now includes SGLang prefill-decode recipes for Zhenwu M890 SuperPoD instances, which cover multi-node tensor and expert parallelism settings. The company also opened invite testing for a new SuperPoD region in Zhongwei, Ningxia. This site includes T-Head's Panmai 920 smart NIC alongside previous high-end configurations.
Alibaba Cloud states the Panmai 920 card uses PCIe 5.0 and 112G PAM4 Ethernet. It reaches 400Gbps throughput and 400 million packets per second, supports multipath RDMA and programmable congestion control, and features an on-card PCIe Switch. The company claims this design allows GPUs and SSDs to connect with lower external-switch dependency, which it says can cut system cost by about 30%.
Day-zero SuperPoD instances for major models are currently available in the Ulanqab and Zhongwei data centers. Alibaba Cloud reports that Lingjun Zhenwu M890 SuperPoD instances are already serving more than 100 large customers for large models, autonomous driving, and embodied AI.
The company also stated that the next-generation Zhenwu V900 SuperPoD instance is expected to deliver about three times the performance of the current generation. This announcement occurred the same week as Huawei Ascend reported its own separate day-zero Kimi K3.1 Pro support, providing two domestic SuperPoD stacks with launch-day paths for the model.
