4 ms·

For reference. The llama.cpp people are not going smaller. Most of those models run on 32 bit floats with the dequantization happening on the fly.