Moonshot AI releases Kimi K3, a 2.8T parameter open-weight MoE model

Moonshot AI has released Kimi K3, an open-weight Mixture of Experts (MoE) model with 2.8 trillion parameters, making it the first in its class to offer public weights. The model utilizes Kimi Delta Attention and a Stable LatentMoE framework, with only a fraction of its parameters active per token for improved efficiency. Kimi K3 is designed for complex tasks like long-horizon coding, agentic workflows, and reasoning, and can be deployed on AWS using Amazon SageMaker HyperPod or Amazon Elastic Kubernetes Service. AI

IMPACT Sets a new benchmark for open-weight MoE models, potentially accelerating self-hosting and research in large-scale AI.

RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on AWS Machine Learning Blog →

• Amazon Elastic Kubernetes Service
• Amazon SageMaker HyperPod
• AWS
• Gated Multi Head Latent Attention
• Hugging Face
• Kimi Delta Attention
• Kimi K3
• Mixture of Experts
• Moonshot AI
• MXFP4
• NVIDIA B300 Blackwell Ultra GPUs
• Stable LatentMoE
• vLLM

• model release
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →