Moonshot AI's Kimi K3 hits 2.8T parameters on AWS
Kimi K3, the first open 2.8T-parameter MoE model, now deployable on AWS with 1M token context
Moonshot AI has officially launched Kimi K3, a 2.8-trillion-parameter open-weight Mixture of Experts (MoE) model that achieves frontier-level intelligence while remaining self-hostable. Built with Kimi Delta Attention (KDA), Gated Multi Head Latent Attention (MLA), and a Stable LatentMoE framework, the model distributes its massive parameter count across 896 specialist experts, activating just 16 per token to deliver 2.5x better scaling efficiency than its predecessor. Kimi K3 supports native multimodal (text + vision) input, tool calling, structured output, and an always-on thinking mode for multi-step problem solving, making it particularly suited for complex reasoning and long-horizon coding tasks.
The model is now deployable on AWS using two approaches: Amazon SageMaker HyperPod and Amazon Elastic Kubernetes Service (EKS). Serving requires a vLLM container (specifically vllm/vllm-openai:kimi-k3) due to its MoE architecture and MXFP4 quantization format. Deployment necessitates an ML.p6-b300.48xlarge instance with 8 NVIDIA B300 Blackwell Ultra GPUs, which AWS provides through Flexible Training Plans or Capacity Blocks. The open weights are available on Hugging Face (moonshotai/Kimi-K3) in MXFP4 format, offering an effective balance between model quality and memory efficiency for large-scale inference.
- Kimi K3 is the first open-weight model to exceed 2.8 trillion parameters, using a Mixture of Experts architecture with 896 experts (16 active per token)
- Requires 8 NVIDIA B300 GPUs (p6-b300.48xlarge instances) for deployment on AWS, available via SageMaker HyperPod or EKS
- Supports native multimodal input, 1M-token context windows, and advanced features like tool calling and structured output
Why It Matters
Kimi K3 enables organizations to self-host a state-of-the-art AI model with unprecedented scale and flexibility, reducing dependency on proprietary APIs.