AWS details two ways to deploy Moonshot AI’s Kimi K3 on AWS infrastructure

AWS describes two deployment approaches for Moonshot AI’s Kimi K3 on its infrastructure: Amazon SageMaker HyperPod and Amazon Elastic Kubernetes Service (Amazon EKS). The post frames these options as solutions for hosting multi-trillion parameter models that require purpose-built infrastructure, high-end GPU compute, and optimized serving frameworks.

Kimi K3 is a 2.8 trillion parameter Mixture of Experts (MoE) model released by Moonshot AI on July 27, 2026, and is described as the first open-weight system to reach the 3 trillion parameter class. The model’s architecture includes Kimi Delta Attention, Gated Multi Head Latent Attention, and a Stable LatentMoE framework. Its 2.8 trillion parameters are distributed across 896 specialist experts, with approximately 16 experts activated per token, yielding about 104 billion active parameters during any single forward pass and a reported 2.5x improvement in scaling efficiency.

The AWS Machine Learning Blog post emphasizes that as open-weight models grow in capability and size, deploying them requires infrastructure tailored for large-scale serving. It positions Amazon SageMaker HyperPod and Amazon EKS as tools to meet these demands, focusing on the operational requirements for running such models in production environments.