FP16
The older 16-bit float: 1 sign bit, 5 exponent bits, 10 mantissa bits. It holds finer detail than BF16 but reaches a much narrower range of magnitudes, so training runs in it can overflow. Both formats are 2 bytes per parameter, so they produce identically sized files.
FP16 came first, from graphics, where the numbers involved are small and well-behaved. Training a large model is neither. BF16 exists because labs kept hitting overflow in FP16 and found they could trade three mantissa bits for the full exponent range and lose nothing that mattered.
If a listing says FP16 or BF16, treat them the same for sizing: 2 bytes a parameter either way.