AI Briefing
KO

Moonshot AI Unveils Kimi K3, a 2.8T-Parameter Model

·2026.07.19 08:30

Key point

Moonshot AI has unveiled Kimi K3, a large-scale open-weight model with 2.8 trillion parameters.

Details

Moonshot AI has unveiled its top-performing model, Kimi K3. This model has a scale of 2.8 trillion (2.8T) parameters, and Moonshot AI introduced it as the world's first open 3T-class model.

Key Technical Features:

  • New Architecture: Combines Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) to optimize information flow.
  • MoE Structure: Uses the Stable LatentMoE framework, which activates only 16 out of 896 experts, to increase computational efficiency.
  • Multimodal and Context: Equipped with native vision capabilities, and supports a long context window of 1 million (1M) tokens.
  • Efficiency: Improved scaling efficiency by approximately 2.5x compared to the previous generation, K2.

Kimi K3 is currently available immediately via API and web service, but the full model weights are scheduled to be released on July 27, 2026. This model is designed on the premise of large-scale infrastructure, and is particularly optimized for long-horizon coding and complex reasoning tasks.

mixture-of-expertsopen-weightsmoonshot-ailarge-language-modelkimi-k3
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Vercel BlogDeepSeek Surpasses Google in Usage; Cost per Token Drops 13.6%RedditMoonshot Releases Kimi K3HuggingFace BlogLiquid AI Releases LFM2.5-VL-DSpark: An Experimental Draft Model for Faster Vision-Language InferencePyTorchKRStepFun Releases Step 5 PreviewTLDR AIFireworks Releases Ember-1: Matches Kimi K3's Performance with 40% Fewer TokensAWS BlogAWS Launches Kimi K3 on Bedrock GA, Releases Corretto 27, and Improves EC2 T8i Performance