Build faster on-device language models through speculative decoding, GPU kernels, quantisation, and tiered memory across consumer hardware.