Alibaba Cloud's Zhenwu M890 Supernode Now Supports Moonshot AI's Kimi K3 Model
Alibaba Cloud announced that its Lingjun Zhenwu M890 supernode instance, powered by T-Head's AI chips, has achieved Day0 adaptation for Moonshot AI's 2.8-trillion-parameter Kimi K3 MoE model. This adaptation leverages joint optimizations across the hardware, software stack, and inference framework to boost inference efficiency. This milestone demonstrates the capability of China's domestic AI hardware and software ecosystem to support massive, multi-trillion-parameter Mixture-of-Experts (MoE) models. It highlights a growing alternative to Nvidia's dominant hardware stack by offering high-speed interconnects and software compatibility. The Zhenwu M890 supernode features the ICN Switch 1.0 interconnect chip, enabling 64 M890 chips to achieve an 800 GB/s All-to-All interconnect speed with 9TB of VRAM. Additionally, T-Head's SAIL software stack allows Kimi's Mooncake inference framework to run out-of-the-box, while native Triton support eliminates the need to rewrite custom operators.
## BACKGROUND
Large language models using the Mixture-of-Experts (MoE) architecture require massive memory and high-bandwidth communication to route tokens to different "expert" sub-networks. T-Head (Pingtouge) is Alibaba's semiconductor division, which recently open-sourced its SAIL software stack to compete with Nvidia's CUDA, while Mooncake is Moonshot AI's specialized, high-performance distributed inference framework.