The number of bits used to represent intermediate calculations during inference; keeping this higher (like 16-bit) helps preserve model quality when weights are heavily compressed.