BF16

Also written: bfloat16

A 16-bit number format that models are commonly trained in. It is the baseline against which FP8 is compared.