Huawei Ascend announces Day-zero inference support for Kimi K3.1 Pro on November 5
Ascend achieved Day-zero inference adaptation for Kimi K3.1 Pro via vLLM Ascend and SGLang, supporting Atlas 800 A3, Atlas 900 A3 SuperPod and Ascend 950 supernode products.
According to the Huawei Compute official account, after Moonshot AI published the full Kimi K3.1 Pro model weights on November 4, Ascend achieved Day-zero inference adaptation through vLLM Ascend and SGLang, supporting inference deployment of Kimi K3.1 Pro on Ascend Atlas 800 A3, Atlas 900 A3 SuperPod supernodes and the Ascend 950 supernode product line. This allows users and developers to begin deployment and application validation on the first day the model ships.
To address the newly added sparse attention, million-scale context and inference service requirements, Ascend carried out joint operator and framework optimization. Kimi K3.1 Pro uses a hybrid linear attention architecture combining KDA linear attention layers and DSA sparse attention layers. Exploiting the computational advantages of this architecture requires coordinated optimization of operator execution, cache management and cross-card communication.
The Ascend inference team said that, drawing on prior adaptation experience with Kimi K3, it conducted model analysis in advance and completed the new sparse attention path and multimodal inference adaptation for K3.1 Pro on top of existing KDA and MoE support, while retaining the existing service capabilities of the inference frameworks. Kimi K3.1 Pro is the first model in the K3.1 series, with 2.8 trillion parameters, native visual understanding and a million token context window.