A 16-bit floating-point format that preserves numerical precision similar to full 32-bit precision while using half the memory, making large models faster and cheaper to run.