[ancientworldnow]

This was trained to be run at FP8 with no quality loss.

[hislaziness]

The model description on huggingface says - Model size - 12.2B params, Tensor type - BF16. Is the Tensor type different from the training param size?