Experimental global target bits‑per‑weight quantization of coder3101/Qwen3.5-2B-heretic
- Using non-standard (forked) LLaMA C++ branch for quantization.
- Using a CLI tool to build KLD evaluation and imatrix calibration datasets for GGUF models, sourced from eaddario/imatrix-calibration.
- Using dataset sources: tools, text_en, text_ru.
- Using dataset chunks: 250.
- Tensors quantinization F16 instead of BF16, Nvidia Pascal architecture friendly like P100.
- Small set of patches added.
- Multimodal work perfectly, use:
Many thanks to Ed Addario for an impressive job.
Quantization comparison
| BPW | PPL correlation | PPL mean ratio | ΔPPL | Mean KLD | Maximum KLD | 99.9% KLD | Mean Δp | RMS Δp |
|---|---|---|---|---|---|---|---|---|
| 5.00 | 99.64% | 1.026057 ± 0.001098 | 0.571999 ± 0.024867 | 0.028004 ± 0.000171 | 2.310003 | 0.489172 | -0.530 ± 0.016 % | 4.045 ± 0.043 % |
| 5.25 | 99.70% | 1.022321 ± 0.001001 | 0.489982 ± 0.022620 | 0.023515 ± 0.000153 | 3.049644 | 0.418944 | -0.442 ± 0.015 % | 3.740 ± 0.044 % |
| 5.30 | 99.70% | 1.021006 ± 0.000993 | 0.461111 ± 0.022363 | 0.022725 ± 0.000146 | 3.486933 | 0.401278 | -0.425 ± 0.015 % | 3.711 ± 0.042 % |
| 5.50 | 99.78% | 1.018126 ± 0.000842 | 0.397896 ± 0.019117 | 0.016384 ± 0.000109 | 2.211204 | 0.266915 | -0.311 ± 0.013 % | 3.141 ± 0.036 % |
| 5.75 | 99.85% | 1.012000 ± 0.000695 | 0.263430 ± 0.015510 | 0.011618 ± 0.000070 | 2.060406 | 0.177300 | -0.278 ± 0.011 % | 2.707 ± 0.029 % |
| 5.80 | 99.82% | 1.015834 ± 0.000776 | 0.347594 ± 0.017563 | 0.013770 ± 0.000105 | 2.561136 | 0.252383 | -0.256 ± 0.012 % | 2.901 ± 0.039 % |
| 6.00 | 99.87% | 1.011607 ± 0.000641 | 0.254793 ± 0.014392 | 0.009732 ± 0.000070 | 2.332554 | 0.151923 | -0.210 ± 0.010 % | 2.452 ± 0.035 % |
| 6.25 | 99.89% | 1.011164 ± 0.000589 | 0.245080 ± 0.013269 | 0.007907 ± 0.000069 | 2.916830 | 0.129800 | -0.203 ± 0.009 % | 2.190 ± 0.033 % |
| 6.30 | 99.90% | 1.011153 ± 0.000577 | 0.244824 ± 0.013017 | 0.007673 ± 0.000058 | 2.056196 | 0.129539 | -0.202 ± 0.009 % | 2.161 ± 0.031 % |
| 6.50 | 99.93% | 1.007030 ± 0.000487 | 0.154329 ± 0.010840 | 0.005530 ± 0.000034 | 0.793070 | 0.086250 | -0.156 ± 0.007 % | 1.828 ± 0.025 % |
| 6.75 | 99.95% | 1.005130 ± 0.000400 | 0.112623 ± 0.008912 | 0.003585 ± 0.000023 | 0.638760 | 0.058130 | -0.082 ± 0.006 % | 1.471 ± 0.020 % |
| 6.80 | 99.95% | 1.004828 ± 0.000397 | 0.105980 ± 0.008823 | 0.003551 ± 0.000022 | 0.470148 | 0.056400 | -0.081 ± 0.006 % | 1.475 ± 0.021 % |
| 7.00 | 99.95% | 1.003980 ± 0.000380 | 0.087374 ± 0.008441 | 0.003187 ± 0.000019 | 0.355283 | 0.045939 | -0.048 ± 0.006 % | 1.381 ± 0.017 % |
| 7.25 | 99.96% | 1.002210 ± 0.000353 | 0.048507 ± 0.007781 | 0.002755 ± 0.000014 | 0.204998 | 0.040443 | -0.056 ± 0.005 % | 1.275 ± 0.013 % |
| 7.30 | 99.96% | 1.002463 ± 0.000351 | 0.054060 ± 0.007734 | 0.002745 ± 0.000016 | 0.321245 | 0.044340 | -0.054 ± 0.005 % | 1.286 ± 0.016 % |
| 7.50 | 99.96% | 1.003666 ± 0.000345 | 0.080484 ± 0.007667 | 0.002463 ± 0.000017 | 0.393500 | 0.044364 | -0.016 ± 0.005 % | 1.201 ± 0.018 % |
| 7.75 | 99.97% | 1.002205 ± 0.000312 | 0.048408 ± 0.006873 | 0.002002 ± 0.000013 | 0.363822 | 0.035939 | -0.038 ± 0.004 % | 1.070 ± 0.017 % |
| 7.80 | 99.97% | 1.001824 ± 0.000313 | 0.040051 ± 0.006892 | 0.001974 ± 0.000012 | 0.194758 | 0.035334 | -0.030 ± 0.004 % | 1.075 ± 0.014 % |
| 8.00 | 99.97% | 1.001736 ± 0.000304 | 0.038111 ± 0.006704 | 0.001844 ± 0.000014 | 0.562781 | 0.031040 | -0.030 ± 0.004 % | 1.031 ± 0.023 % |
| 8.25 | 99.97% | 1.001911 ± 0.000288 | 0.041960 ± 0.006350 | 0.001549 ± 0.000012 | 0.438731 | 0.028080 | -0.017 ± 0.004 % | 0.948 ± 0.020 % |
| 8.30 | 99.98% | 1.001561 ± 0.000280 | 0.034274 ± 0.006185 | 0.001485 ± 0.000010 | 0.330964 | 0.026834 | -0.011 ± 0.004 % | 0.923 ± 0.017 % |
| 8.50 | 99.98% | 1.001301 ± 0.000265 | 0.028563 ± 0.005830 | 0.001258 ± 0.000018 | 0.926594 | 0.022297 | -0.016 ± 0.004 % | 0.885 ± 0.036 % |
| 8.75 | 99.98% | 1.002038 ± 0.000232 | 0.044738 ± 0.005156 | 0.000886 ± 0.000006 | 0.168904 | 0.012202 | -0.020 ± 0.003 % | 0.713 ± 0.012 % |
| 8.80 | 99.98% | 1.002025 ± 0.000232 | 0.044443 ± 0.005160 | 0.000881 ± 0.000006 | 0.183456 | 0.011982 | -0.021 ± 0.003 % | 0.715 ± 0.013 % |
| 9.00 | 99.98% | 1.001946 ± 0.000227 | 0.042725 ± 0.005036 | 0.000830 ± 0.000004 | 0.080291 | 0.010794 | -0.013 ± 0.003 % | 0.672 ± 0.007 % |
| 9.25 | 99.98% | 1.001806 ± 0.000225 | 0.039656 ± 0.004985 | 0.000805 ± 0.000004 | 0.133142 | 0.010662 | -0.012 ± 0.003 % | 0.669 ± 0.009 % |
| 9.30 | 99.98% | 1.001752 ± 0.000225 | 0.038468 ± 0.004984 | 0.000792 ± 0.000004 | 0.100964 | 0.010653 | -0.012 ± 0.003 % | 0.676 ± 0.009 % |
| 9.50 | 99.98% | 1.001729 ± 0.000221 | 0.037948 ± 0.004912 | 0.000770 ± 0.000004 | 0.070395 | 0.009889 | -0.010 ± 0.003 % | 0.659 ± 0.008 % |
| 9.75 | 99.99% | 1.001773 ± 0.000218 | 0.038927 ± 0.004847 | 0.000745 ± 0.000004 | 0.070781 | 0.010044 | -0.012 ± 0.003 % | 0.649 ± 0.008 % |
| 9.80 | 99.99% | 1.001839 ± 0.000217 | 0.040373 ± 0.004831 | 0.000741 ± 0.000004 | 0.144547 | 0.008833 | -0.013 ± 0.003 % | 0.652 ± 0.010 % |
| 10.00 | 99.99% | 1.001912 ± 0.000214 | 0.041973 ± 0.004767 | 0.000704 ± 0.000004 | 0.142314 | 0.008692 | -0.011 ± 0.003 % | 0.637 ± 0.010 % |
| 10.25 | 99.99% | 1.001895 ± 0.000212 | 0.041595 ± 0.004703 | 0.000671 ± 0.000005 | 0.192968 | 0.009046 | -0.010 ± 0.003 % | 0.631 ± 0.012 % |
| 10.30 | 99.99% | 1.001793 ± 0.000211 | 0.039367 ± 0.004686 | 0.000663 ± 0.000004 | 0.080132 | 0.008241 | -0.011 ± 0.002 % | 0.614 ± 0.007 % |
| 10.50 | 99.99% | 1.001711 ± 0.000208 | 0.037563 ± 0.004625 | 0.000647 ± 0.000003 | 0.047110 | 0.008988 | -0.009 ± 0.002 % | 0.611 ± 0.007 % |
| 10.75 | 99.99% | 1.001667 ± 0.000204 | 0.036600 ± 0.004524 | 0.000602 ± 0.000003 | 0.055274 | 0.008006 | -0.011 ± 0.002 % | 0.585 ± 0.007 % |
| 10.80 | 99.99% | 1.001796 ± 0.000204 | 0.039420 ± 0.004538 | 0.000601 ± 0.000003 | 0.043833 | 0.007718 | -0.017 ± 0.002 % | 0.580 ± 0.006 % |
| 11.00 | 99.99% | 1.001853 ± 0.000202 | 0.040683 ± 0.004499 | 0.000589 ± 0.000003 | 0.076128 | 0.007433 | -0.014 ± 0.002 % | 0.572 ± 0.006 % |
| 11.25 | 99.99% | 1.001751 ± 0.000201 | 0.038444 ± 0.004463 | 0.000568 ± 0.000003 | 0.052181 | 0.007172 | -0.011 ± 0.002 % | 0.563 ± 0.005 % |
| 11.30 | 99.99% | 1.001664 ± 0.000201 | 0.036527 ± 0.004456 | 0.000563 ± 0.000003 | 0.078151 | 0.006822 | -0.008 ± 0.002 % | 0.566 ± 0.006 % |
| 11.50 | 99.99% | 1.001737 ± 0.000199 | 0.038121 ± 0.004414 | 0.000550 ± 0.000003 | 0.057737 | 0.006762 | -0.010 ± 0.002 % | 0.552 ± 0.006 % |
| 11.75 | 99.99% | 1.001612 ± 0.000198 | 0.035386 ± 0.004396 | 0.000536 ± 0.000003 | 0.061081 | 0.006804 | -0.009 ± 0.002 % | 0.547 ± 0.005 % |
| 11.80 | 99.99% | 1.001622 ± 0.000196 | 0.035598 ± 0.004363 | 0.000530 ± 0.000003 | 0.063305 | 0.006763 | -0.008 ± 0.002 % | 0.542 ± 0.006 % |
- Downloads last month
- 16,032
Hardware compatibility
Log In to add your hardware
We're not able to determine the quantization variants.