Mixture-of-experts scaling

Sparse architectures trading total parameter count against active parameters per token across flagship and mid-size models.