Following Moonshot AI's release of the full model weights for Kimi K3.1 Pro, Huawei Ascend completed 0-day inference adaptation the very same day, enabling direct deployment validation. The model is a 2.8-trillion-parameter MoE (Mixture of Experts) model with native support for visual understanding and a massive one-million-token context window, delivering significant improvements over the previous generation in tasks such as code writing, 3D creation, financial analysis, and long-document processing. It adopts a hybrid linear attention architecture and adds a sparse attention module, which requires targeted optimization of operators, caching, and multi-card communication. This adaptation covers the Atlas 800 A3, Atlas 900 A3 SuperPod, and Ascend 950 super-node hardware, and is implemented through two inference frameworks: vLLM Ascend and SGLang. Drawing on prior experience with Kimi K3 adaptation, the Ascend team analyzed the model structure in advance and completed adaptation for the newly added sparse attention and multimodal inference. Through distributed caching, the cache is spread across multiple NPU cards to ease memory pressure on any single card, and core operators were modified to eliminate redundant data copying and wasted computation, thereby improving hardware utilization. Deployment supports direct loading of native weights, and can also enable module-specific differentiated low-bit quantization combined with attention cache quantization to reduce memory footprint and conserve hardware resources. Existing capabilities such as prefix caching, chunked prefill, and speculative inference are all retained, allowing developers to integrate the model into existing inference services without changing their usage habits. Official deployment documentation for both frameworks has been published for reference.