The new Chinese LLM is not what (some) people think.
Kimi K3, the new open-weights LLM from Chinese lab Moonshot AI, dropped last week. It’s a powerful model that, perhaps inevitably, sparked debate — including accusations of ushering in decelerationist ‘AI communism’ (whatever that means).
Amidst the noise, a number of AI commentators (including the Hard Fork guys and Scott Galloway) have asserted that K3 is noteworthy because it’s nearly as good as US models, at a fraction of the cost.
Unfortunately, they are wrong. Just because a model is Chinese, it is not magically cheap. And in fact, K3 isn’t cheap. Across Artificial Analysis’s ‘Intelligence Index’ benchmark, it cost slightly less per task than OpenAI’s top model. Compared to other Chinese models, it’s positively expensive - double the cost of GLM-5.2, and ~20x more than DeepSeek V4.
There’s nuance, of course. Kimi costs less for coding, but more for office work. It’s verbose, so it takes longer to solve problems. It’s open-weights, so you could (with a few GPUs) finetune it for your own purposes. But we’re not talking nuance here, this is just about one claim that I’ve seen/heard enough to find slightly annoying.
In summary: Is K3 a big step forwards for Chinese LLMs (and the open-weights people)? Yes. Is it concerning for the US market that a Chinese lab have fast-followed OpenAI and Anthropic with much less capex, possibly aided by distillation? Also yes. Is it cheap? No. No it isn’t. Please stop saying that. And please stop saying it’s another ‘DeepSeek moment’ too.
Posted on July 26, 2026