[model performance] apply quantization

[dmytro-omelian]

Reduces the precision of the weights from float32 to int8, which can decrease model size and increase inference speed with a slight cost to accuracy.