Moonshot AI releases model weights for Kimi K3.1 Pro

Moonshot AI has released the weights for its Kimi K3.1 Pro model, allowing developers to deploy the system independently. The model utilizes a mixture-of-experts architecture consisting of 2.8 trillion parameters, with approximately 104 billion parameters active during inference. It is designed to handle a context window of approximately one million tokens.

This release expands upon the capabilities of the Kimi K3 model by incorporating further reinforcement learning specifically targeted at coding, 3D asset creation, financial research, and long-running autonomous agent tasks. Previously, access to the model was restricted to Moonshot AI’s own application interface and API.

Furthermore, Huawei has confirmed inference support for the model through its vLLM Ascend and SGLang integrations. This compatibility covers the Atlas A3 and Ascend 950 hardware systems, facilitating the use of the model within enterprise infrastructure.