(computing) A number format representing a real number consisting of a sign bit, a variable-size "regime" part (which modifies the exponent), up to two exponent bits, and a fraction part, proposed as a more efficient alternative to IEEE 754 floats in AI applications.