A 16-bit numerical format that balances memory efficiency with numerical stability, using fewer bits than standard 32-bit floats while maintaining training and inference quality.