A low-precision floating-point format (4-bit) designed for efficient neural network computation while maintaining reasonable accuracy.