Is it though? I got sub 10ms for 200M params vision model way back in 2020 with TensorRT optimization running on RTX 2080Ti. 30ms for 0.8B classifier model doesn't sound special to me.