A compressed number format that uses 8 bits instead of the standard 32 bits, dramatically shrinking model size at the cost of slightly reduced precision.