Overview

Kimi K3 is Moonshot’s most capable model and one of the largest open-weight models released. It is built on Kimi Delta Attention (KDA) and Attention Residuals, activates 16 of 896 experts per token, and understands text, images and video. It launched in Moonshot’s apps and API in July 2026, with weights on Hugging Face later that month.

Highlights

  • 2.8T total, 104B active parameters
  • Native text, image and video understanding
  • One-million-token context window
  • Stable LatentMoE framework, about 2.5x more scaling-efficient than Kimi K2 per Moonshot

Typical uses

Long-horizon coding, agentic knowledge work and multimodal research, mainly via the API.

Things to consider

  • The weights are over a terabyte; local use is impractical for almost everyone.
  • Released under the custom Kimi K3 License, not the Modified MIT of K2.

Specifications are taken from the model card on Hugging Face and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.