Alibaba Cloud's Lingjun Zhenwu M890 node supports Kimi K3 with 2.8 trillion parameters

Alibaba Cloud says its Lingjun Zhenwu M890 super node now supports Kimi K3, a model with 2.8 trillion parameters from Moonshot AI. The model uses a mixture-of-experts (MoE) design and is one of the largest models around. The M890 node is built around Pingtouge's Zhenwu M890 AI chip, and uses ICN Switch 1.0 to connect 64 chips at 800 GB/s, with 9TB of memory. This setup speeds up communication for MoE models and boosts token generation. Alibaba Cloud, Pingtouge, and the Kimi team worked closely on chips and software, so Kimi K3 runs smoothly from day one. The T-Head SAIL software stack lets Kimi's Mooncake inference framework work right away, and M890's Triton support means custom operators don't need to be rewritten, cutting down adaptation work.