Huawei's Ascend computing team has completed day-zero inference adaptation for Kimi K3.1 Pro, a 2.8-trillion-parameter model. According to the Huawei Computing official account, as reported by IT Home and ifeng Tech, the model's full weights were released by Moonshot AI on November 4.
Developers can deploy the model on Atlas 800 A3 servers, the Atlas 900 A3 SuperPoD, and the Ascend 950 SuperPoD series starting from the day of release using vLLM Ascend and SGLang. Kimi K3.1 Pro is the first model in the K3.1 series and is a mixture-of-experts model featuring native visual understanding and a context window of about one million tokens. The model activates roughly 104 billion parameters per inference step.
The architecture of the model, which includes K3's hybrid linear attention built on Kimi Delta Attention (KDA) layers and new DSA sparse attention layers, required the most adaptation work. Huawei stated that its inference team built on previous Kimi K3 adaptation by analyzing the new model in advance and performing model integration, operator development, and service validation in parallel. Because KDA and MoE support were already in place, the team added the new sparse attention path and multimodal inference while maintaining existing serving features.
Huawei reported that the optimization covered core operators for residual computation, preprocessing, sparse attention computation, and historical information selection. These were combined with framework-level features and distributed caching to support production serving and long-context inference. Huawei stated the goal was to allow users and developers to begin deployment and application testing on the first day a model ships.
The vLLM Ascend documentation lists Kimi K3.1 Pro as experimental support in vLLM Ascend 0.30.0. Reference designs for the model use eight-node prefill-decode separation, consisting of four prefill and four decode nodes, including a one-million-token long-context configuration. While Atlas 800I A3 servers run a W4A8 quantized checkpoint, 950DT products use a W4A8 MXFP variant. The guide specifies that Atlas A2 and 950PR products are not yet supported.
The model's size presents practical hardware requirements. IT Home noted that the BF16 weights occupy about 5.6TB and require approximately 1.4TB with INT4 quantization, which rules out single consumer GPUs. Huawei has positioned its A3 and 950 SuperPoD systems for this class of trillion-parameter MoE inference, and Kimi K3.1 Pro is the latest open-weight model the company has supported on launch day.
