Model
Inference & Serving
Kimi K2
1Moonshot 2AI's 3trillion-parameter 4open 5MoE 6model
the six Ws · specification
W1
Who
Moonshot AI, a Chinese AI lab, developed the Kimi K2 model series.
W2
What
Kimi K2 is a mixture-of-experts LLM with 1 trillion total and 32 billion activated parameters, tuned for agentic and coding tasks.
W3
Where
Weights are hosted on Hugging Face under moonshotai, with code at github.com/MoonshotAI/Kimi-K2.
W4
When
Released in July 2025.
W5
Why
Built to deliver frontier-level agentic and coding performance as an openly licensed alternative to closed frontier models.
W6
With
Runs via vLLM or SGLang; requires multi-GPU or multi-node inference infrastructure given its scale.
W7
Watch
Released under a Modified MIT License, weights open, no credentials required to download, self-hosted or via Moonshot's API; maintained by Moonshot AI.
alternatives
works with