I guess the next step would be adding support for quantized arithmetic.
And the lower precision float variants.
And the lower precision float variants.