I still wanna know how Luna is so token efficient compared to its peers (dsv4-flash, glm5.3-flash, etc.).

It uses like 40% of the tokens for comparable results.

How do they do it? It's not a way bigger model. It's some kinda posttraining trick. How?