Huawei Ascend Announces 0 Day Support for Kimi K3.1 Pro MoE Model
Huawei announced '0 Day' Ascend hardware and framework support for the inference deployment of Moonshot AI's 2.8-trillion parameter Kimi K3.1 Pro MoE model. The integration was simultaneously achieved through vLLM Ascend and SGLang to support products like Atlas 800 A3 and Atlas 900 A3 SuperPod. This development demonstrates strong hardware-software co-optimization between domestic Chinese AI hardware and advanced large-scale multimodal models. It allows users and developers to deploy and validate massive 2.8-trillion parameter models on launch day. The K3.1 Pro model features a mixed linear attention architecture combining KDA and DSA sparse attention layers, requiring specialized operator execution and cross-card communication optimizations. Huawei's inference team completed the sparse attention path and multimodal inference adaptation leveraging previous Kimi K3 experience.
## BACKGROUND
Huawei Ascend is an AI computing infrastructure ecosystem built around Ascend processors, featuring the CANN heterogeneous computing architecture, development toolchains, and hardware products like Atlas servers and clusters. Mixture of Experts (MoE) is a machine learning technique that divides a task space into multiple sub-tasks handled by specialized sub-networks, enabling efficient scaling of model parameters.