It uses 75% linear attention layers so it is inherently lower cost. And it is MOE so active parameters are far lower.