It will be interesting to see the token efficiency analysis. This is my first question now with Chinese models; I take raw benchmark performance for granted.