Vector-quantized gemma-4 for MLX: 26B indistinguishable from bf16 at 18.7G; e4b with a better embedding table at 7.4G.