Huawei ascend platform enables deployment of kimi K3.1 pro, advanced moe model

Huawei announced that its Ascend platform now supports the inference deployment of Kimi K3.1 Pro, a model featuring 2.8 trillion parameters and advanced Mixture-of-Experts (MoE) capabilities. This support was achieved by integrating Ascend with vLLM and SGLang to enable 0 Day inference adaptation, compatible with Ascend Atlas 800 A3, Atlas 900 A3 SuperPod, and Ascend 950 SuperPod series products. Kimi K3.1 Pro, the first model in the K3.1 series, boasts native visual understanding and a million-token context window. The model's enhanced capabilities in areas like code development, 3D creation, financial research, and long-term agent tasks stem from an expanded post-training scale over the K3 base model. To support the model's requirements, such as sparse attention and massive context handling, Ascend conducted collaborative optimizations across its operators and frameworks. The inference team leveraged prior Kimi K3 adaptation experience to integrate the new sparse attention paths and multimodal inference support, covering core optimizations like historical information filtering and residual calculation.