# Ornith-1.0-9B — quantization comparison (wikitext-2 test) Reference: **BF16**, PPL = 8.1268. KLD = KL-divergence of each quant's output distribution vs. BF16 (lower = closer to full precision; Top-1 = how often the argmax token matches BF16). | Scheme | Size (GB) | BPW | PPL | ΔPPL % | Mean KLD | KLD q99 | Top-1 agree % | |--------|----------:|----:|----:|-------:|---------:|--------:|--------------:| | BF16 (ref) | — | 16.00 | 8.1268 | 0.00 | 0.000000 | 0.000000 | 100.00 | | Q2_K_S | 3.70 | 3.30 | 9.8933 | +21.74 | 0.295998 | 2.7466 | 76.83 | | Q3_K_S | 4.26 | 3.81 | 8.5273 | +4.93 | 0.121556 | 1.0865 | 85.00 | | Q3_K_M | 4.52 | 4.04 | 8.5692 | +5.44 | 0.121405 | 1.0682 | 84.87 | | Q3_K_L | 4.93 | 4.40 | 8.6009 | +5.83 | 0.120064 | 1.0568 | 84.90 | | Q4_K_S | 5.35 | 4.78 | 8.0475 | -0.98 | 0.042472 | 0.3598 | 91.01 | | Q4_K_M | 5.63 | 5.03 | 8.0752 | -0.63 | 0.040609 | 0.3486 | 91.21 | | Q6_K | 7.36 | 6.58 | 8.1500 | +0.29 | 0.004020 | 0.0252 | 96.64 | | Q8_0 | 9.53 | 8.51 | 8.1361 | +0.12 | 0.002044 | 0.0125 | 97.51 |