A Beijing-based startup just dropped what it calls the largest open-weight multimodal AI model ever built. Moonshot AI unveiled Kimi K3 on July 16, packing 2.8 trillion parameters into a system that processes text, images, and video with a context window stretching to one million tokens.
What Kimi K3 actually does
Kimi K3 uses architectural tricks called Kimi Delta Attention and Attention Residuals that activate only a small fraction of those parameters for each token processed. The model is massive but efficient, like a warehouse full of specialists where only the relevant ones show up to work on any given task.
Kimi K3 currently leads the Frontend Code Arena with 1,679 points. Across broader AI leaderboards, it ranks second or third, which means it hasn’t dethroned OpenAI’s or Anthropic’s top models across the board.
Moonshot AI is pricing API access at $3 per million input tokens and $15 per million output tokens. Full open weights are expected by July 27, which would make Kimi K3 the largest openly available model of its kind.
