bf16 is generally easier to train neural network than fp16 on due to no need for scaling. And most model training and inference performs the same with fp32 and bf16.