[gruez]

>so they place much more emphasis on compute-efficiency.

Maybe on training, but on inference they use more tokens than comparable western models.

https://artificialanalysis.ai/?output-tokens=intelligence-vs...