They cannot be copies, because all of them have different structures, which are explained in published research papers, while nobody knows the structure of the American SOTA models.
It is also impossible for China to have accessed the hundreds or thousands of terabytes of data that must have been hoarded by OpenAI, Anthropic or Google, so the Chinese must have gathered training data sets by themselves, by crawling the Internet and using pirate libraries, the same like Anthropic and OpenAI.
If the American allegations were true, then some Chinese companies might have extracted some blackbox data about the American models, by API querying, in an amount small enough that it could be used only for the post-training of a model created independently.
Such an information extraction is a breach of contract, but it cannot be called "copying", because it can be only an infinitesimal fraction of the information stored in a SOTA LLM.
Europe is actually catching up right now. Not at the frontier like kimi and qwen 3.6 (yet) but it's just a matter of time until the European A3B models catch up, too.
Apertus and Soofi, both open source (not only open weight) models:
The Chinese open weights models are copies of the American closed ones. Nobody in the US is trying to do that, I guess.
They cannot be copies, because all of them have different structures, which are explained in published research papers, while nobody knows the structure of the American SOTA models.
It is also impossible for China to have accessed the hundreds or thousands of terabytes of data that must have been hoarded by OpenAI, Anthropic or Google, so the Chinese must have gathered training data sets by themselves, by crawling the Internet and using pirate libraries, the same like Anthropic and OpenAI.
If the American allegations were true, then some Chinese companies might have extracted some blackbox data about the American models, by API querying, in an amount small enough that it could be used only for the post-training of a model created independently.
Such an information extraction is a breach of contract, but it cannot be called "copying", because it can be only an infinitesimal fraction of the information stored in a SOTA LLM.
> Nobody in the US is trying to do that, I guess.
Yes, but WHY, I wonder.
Seems like a business opportunity.
Europe is actually catching up right now. Not at the frontier like kimi and qwen 3.6 (yet) but it's just a matter of time until the European A3B models catch up, too.
Apertus and Soofi, both open source (not only open weight) models:
[1] https://apertus.ai/en/apps/apertus-model/
[2] https://www.soofi.info/