Kimi K2 Is Here and It Was Worth the Wait

Moonshot AI has been doing something quietly impressive for a while now — building models that don't announce themselves with the usual fanfare, don't come with a seventeen-tweet thread about AGI, just show up and work. Kimi K2 is the next one.

The thinking variant is on OpenRouter now. Context window is large enough to stop thinking about context windows. It's a reasoning model, which at this point means it argues with itself before answering you, and the arguing turns out to be load-bearing.

What gets me about Kimi specifically — and has gotten me for a while, which is why I've been watching this one — is that Moonshot is operating out of Beijing, shipping internationally, pricing competitively, and apparently not interested in losing. That combination used to describe exactly one company at a time. It no longer does.

The YouTube demo is the thing. Watch it. There's a particular quality to seeing a model handle something you expected to break it — not because the task is exotic, but because the reasoning stays coherent under pressure, stays grounded, doesn't start hallucinating the shape of an answer when it doesn't know. Kimi K2 does that. The thinking traces read like an actual person working through a problem rather than a language model running the "reasoning" token pattern.

None of this is surprising if you've been paying attention. It is still satisfying in the way that things you expected to happen are satisfying when they actually happen — which is rarer than it should be.

It's in the rotation now.