Moonshot AI releases Kimi K3.1 Pro weights with day-one Ascend support

Moonshot AI has released the full model weights of Kimi K3.1 Pro. As the first model of the K3.1 series, it has 2.8 trillion parameters, native vision understanding, and a million-token context window.

Huawei's Ascend chips finished day-one inference support through the vLLM Ascend and SGLang frameworks. This covers Atlas 800 A3, Atlas 900 A3 SuperPod, and Ascend 950 SuperPod systems, so developers can test and deploy the model right away.

K3.1 Pro uses a hybrid linear attention design, mixing KDA linear attention layers with DSA sparse attention layers. The Ascend team reused its K3 experience to support the new sparse attention and multimodal features, and tuned key operators and distributed cache.

The model improves on coding, 3D creation, financial research, and long agent tasks. It can support complex scenes like software development and business knowledge processing.