naklitechie's picture
Add q4f16 variants (810 MB total, 64% smaller than fp16). Browser-ready for WebGPU MatMulNBits.
d4598da verified