test Model Size SSIM Similality 推奨 FP32 19 GB 100.0 % △ FP16 9.6 GB 98.0 % 〇 FP8 4.8 GB 95.3 % △ Q8_0 5.1 GB 97.6 % 〇 Q6_K 4.0 GB 97.3 % △ Q5_K_M 3.4 GB 94.8 % Q4_K_M 2.9 GB 96.4 %