在 AWS 上部署 Kimi K3：SageMaker HyperPod 与 EKS 实践
原标题：Deploying Kimi K3 on Amazon SageMaker HyperPod and Amazon EKS
AI 摘要
AWS 机器学习博客发布文章，介绍如何在 Amazon SageMaker HyperPod 和 Amazon EKS 上部署 Moonshot AI 于 2026 年 7 月 27 日发布的 2.8 万亿参数 MoE 模型 Kimi K3。该模型采用 Kimi Delta Attention 和 Gated MLA 等架构，激活参数 1040 亿，支持原生多模态和 100 万上下文。文章提供了两种部署方案，并强调了 p6-b300 实例和 vLLM 推理容器的要求。
正文节选
Deploying Kimi K3 on Amazon SageMaker HyperPod and Amazon EKS Open weight models have become powerful enough to handle complex tasks such as multi-step agentic workflows, advanced reasoning, and long-horizon coding. However, as these models grow in capability, they also grow in size and hosting multi-trillion parameter architectures requires purpose-built infrastructure, high-end GPU compute, and optimized serving frameworks. On July 27, 2026, Moonshot AI released Kimi K3, a 2.8 trillion paramet