> Isn't there pretty good indications that the chinese llms have been trained on top of the expensive models?

So what ?

> Isn't there pretty good indications that the chinese llms have been trained on top of the expensive models?

there are pretty good indications that the american llms have been trained on top of stolen data

> Their cost is not real.

They can’t even officially account for any nvidia gpus they managed to buy outside the official channels.

This so-called "PC compatible" seems like a cheap copy, give me a real IBM every time.

There are many providers (Fireworks, Groq, Cerebras, Google Vertex), some using rather common hardware from Nvidia, etc., others their own solutions focused solely on high throughput inference. They often tend to be faster, cheaper and/or more reliable than what the lab that trained the model is charging [0], simply because there is some competition, unlike with US frontier models which at best can be hosted by Azure, AWS or GCloud at the same price as the first party.

[0] https://openrouter.ai/moonshotai/kimi-k2-thinking

I'm still testing myself and cannot make a confident statement yet, but Artifical Analysis is a solid and independent source for a general overview: https://artificialanalysis.ai/

Kimi K2.5 is rather competitive in regard to pure output quality, agentic evals are also close to or beating US made frontier models and, lest we forget, the model is far more affordable than said competitors, to a point where it is frankly silly that we are actually comparing them.

For what it's worth, of the models I have been able to test as of yet, many purely on performance (meaning solely task adherence, output quality and agentic capabilities; so discounting price, speed, hosting flexibility), I have personally found the prior Kimi K2 Thinking model to be overall more usable and reliable than Gemini 3 Pro and Flash. Purely on output quality in very specific coding tasks, Opus 4.5 was in my testing leaps and bounds superior of both the Gemini models and K2 Thinking however, though task adherence was surprisingly less reliable than Haiku 4.5 or K2 Thinking.

Being many times more expensive and in some cases less reliably adhering to tasks, I really cannot say that Opus 4.5 is superior or Kimi K2 Thinking is inferior here. The latter is certainly better in my specific usage than any Gemini model and again, I haven't yet gone through this with K2.5. I try not to just presume from the outset that K2.5 is better than K2 Thinking, though even if K2.5 remains at the same level of quality and reliability, just with multi modal input, that'd make the model very competitive.

Open models have been about 6 to 9 months behind frontier models, and this has been the case for the last 2 years. That is a very long time for this technology at it's current state of development.