Moonshot AI Releases Kimi K3, a 2.8 Trillion Parameter Open-Weight Model; AWS Publishes Deployment Guide

What happened

Moonshot AI released Kimi K3 on July 27, 2026, an open-weight Mixture-of-Experts (MoE) model with 2.8 trillion total parameters — according to AWS, the first open-weight system to reach the 3 trillion parameter class. AWS has since published a deployment guide for running the model on its infrastructure via Amazon SageMaker HyperPod and Amazon EKS.

Architecture and specs

Kimi K3 distributes its 2.8 trillion parameters across 896 specialist experts, activating only 16 per token. That yields roughly 104 billion active parameters per forward pass — a design Moonshot AI claims delivers a 2.5x improvement in scaling efficiency over its predecessor, Kimi K2.

The model combines three architectural components: Kimi Delta Attention (KDA), Gated Multi-Head Latent Attention (MLA), and a framework Moonshot AI calls Stable LatentMoE. It supports a 1 million token context window and is natively multimodal, handling both text and vision inputs. Moonshot AI says the model excels at long-horizon coding, agentic workflows, and complex reasoning, with native tool calling, structured output, and an always-on "thinking mode" for multi-step problem solving.

Weights are published on Hugging Face under moonshotai/Kimi-K3, distributed in MXFP4 (4-bit microscaling floating point) format to reduce the memory footprint required for inference at this scale. Pricing for API access has not been disclosed; the weights are open and self-hostable, though the underlying compute cost is substantial.

Serving requirements

Because vLLM's mainline container does not yet support the model's MoE and MXFP4 configuration, AWS notes that serving currently requires a day-0 container build (vllm/vllm-openai:kimi-k3) pending merge into vLLM's main release. AWS specifies that deployment requires an ml.p6-b300.48xlarge instance — 8 NVIDIA B300 Blackwell Ultra GPUs with high-bandwidth interconnects — to run tensor-parallel inference across the model's full expert pool.

Deployment paths on AWS

AWS outlines two deployment routes:

• SageMaker HyperPod: Cluster creation with EKS orchestration and the Inference Operator, which automates container scheduling, model download, and endpoint health checks. Capacity for B300 instances is procured through Flexible Training Plans, a committed reservation mechanism.
• Amazon EKS (self-managed): Using the AI on EKS project's Terraform-based cluster recipes, with capacity reserved through EC2 Capacity Blocks for a fixed duration without long-term commitment.

Both paths expose an OpenAI-compatible v1/chat/completions endpoint once the deployment is active.

What this means

Kimi K3's release pushes the open-weight frontier past the 3-trillion-parameter mark, but its size makes self-hosting a serious infrastructure undertaking — a single instance requires 8 Blackwell Ultra GPUs, and reserved capacity (not on-demand) is effectively mandatory given current GPU supply constraints. AWS's same-day deployment guide signals the company is positioning HyperPod and EKS as the default enterprise path for running frontier open-weight models, ahead of upstream tooling like vLLM formally supporting the architecture. For most organizations, Kimi K3 will remain accessible primarily through hosted API providers rather than self-managed deployment, given the hardware and operational overhead AWS's own guide makes explicit.