“On some benchmarks, the MuseSpark model achieves results similar to competitor models while using significantly fewer tokens.”