This is mostly about inference speed, while maintaining long context performance.