It implies the IEEE 754R 128-bit float, but in practice is typically whatever long double is on the platform, which #10281 shows can sometimes be other types.