Lingjun Zhenwu M890 supports Kimi K3.1 Pro: beta testing opens in Zhongwei
On November 5, Alibaba Cloud announced day-0 support for the model weights of Moonshot AI's Kimi K3.1 Pro. Alibaba Cloud has completed full end-to-end validation on two leading open-source inference frameworks, vLLM and SGLang, enabling users to rapidly deploy the model on Alibaba Cloud Supernode instances. Invitation-only beta testing for Supernodes is also now open in the China (Zhongwei) region to further increase the availability of large-scale computing power.
Kimi K3.1 Pro is the latest open-source large language model from Moonshot AI, designed for on-premises or cloud deployment. It officially features 2.8 trillion parameters, making it one of the largest open-source models globally, and supports a 1 million-token context window. The model continues to use a massive-scale sparse Mixture-of-Experts (MoE) architecture and employs a hybrid attention design that combines KDA linear attention with DSA sparse attention. It also supports Multi-Token Prediction (MTP). This architecture places extremely high demands on memory capacity and high-speed scale-up interconnects.
To meet the high-throughput, low-latency, and high-concurrency demands of trillion-parameter models like K3.1 Pro, Alibaba Cloud, T-Head, and Moonshot AI collaborated on deep adaptations using the T-Head SAIL software stack. Key optimizations focused on the adaptation and tuning of FlashMLA operators, DeepGEMM MOE operators, MegaMoE, and MoonEP. You can rapidly deploy inference services on Supernode resources using either bare metal or the REIN intelligent engine service.
As the first large-scale, commercially available cloud Supernode in China, Alibaba Cloud's Lingjun Zhenwu Supernode instances enable customers to quickly and reliably host large-scale, flagship open-source models while innovating to meet evolving market demands. The Supernode instances in the China (Zhongwei) region are also now available for invitation-only beta testing. In addition to the existing 9 TB GPU memory version, a new 6 TB GPU memory Supernode instance is now available. These instances are built on the T-Head Panmai 920 intelligent network card, which uses PCIe 5.0 and 112G PAM4 Ethernet technology. The card supports up to 400 Gbps of throughput, a packet processing rate of 400 Mpps, and multipath RDMA with user-programmable congestion control. Its built-in PCIe Switch chip-level network architecture directly connects the GPU and SSD with extremely low latency, reduces dependency on external switch chips, and cuts system costs by up to 30%. Alibaba Cloud now offers Supernode instances with day-0 support for various mainstream large models in both its Ulanqab and Zhongwei data centers.
Currently, Lingjun Zhenwu M890 Supernode instances serve hundreds of leading customers across various industries, including large model development, autonomous driving, and embodied intelligence. The upcoming next-generation Zhenwu V900 Supernode instance will deliver 3x the performance of the current generation. Alibaba Cloud and T-Head will leverage their large-scale clusters and mature hardware and software acceleration ecosystem to continue working with customers to accelerate the development of foundational models and advance industrial intelligence.