The thing that is interesting though is that China is releasing the models they train (for the most part). Unlike OpenAI, for instance, China seems to still be developing in an open way including sharing code, weights, and publishing techniques.

These models are pretty amazing too. Their performance remains high even with smaller models, allowing them to competitively run on consumer grade hardware. I can (and do) run the Qwen models at home and use them for real tasks. As the tool ecosystem matures we can offload what the model is bad at to other systems, which makes it even easier to use smaller models. The focus on efficiency and performance from the Chinese companies has had huge practical gains.

Any US startup, or even large company, in the AI space can and should take advantage of this and use these models. The cost benefit compared to something like Azure or OpenAI is massive.