Hypothesis
At any given file size, Agention Precision tiers sit closer to the BF16 reference than every other published quantization of Qwen3.8-27B, and the ordering holds on all three corpora rather than on a single favourable one.
Shared results · Agention Precision tiers, Qwen3.8-27B
Cross-vendor fidelity at matched bytes
Published 25 September 2026 · 25 finished runs · 3 charts
- hardware
- strix-halo-vulkan
- llamacpp build
- 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47
Hypothesis
Charts
Fidelity vs file size, held-out prose
Colour: packager.
Frontier chosen by hand: 8 runs, joined in file size (GiB) order. The dashed line connects recorded runs; it does not imply results between them.
How this chart is drawn
One point per successful run, no aggregation. Context series separate experiments, recorded comparison context, input references and environments; matching metadata does not establish experimental equivalence.
Schema versions contributing: quant-fidelity v1: 0 runs · quant-fidelity v2 (pinned): 26 runs. Versions join only where every charted field still records the same kind of value in the same unit and direction.
Displayed Y range: 0.01114 to 0.2793 · X range: 8.17 to 16.35
Schema 3dac891d-f938-47d2-84fc-304eb64adb94 · 25 eligible runs · 1 excluded · Live as of 2026-09-25T17:47:39.289Z
Source data and excluded runs
| Run | Experiment | Revision | file size (GiB) | KL divergence vs BF16 (lower is better) | Series | Context | Frontier |
|---|---|---|---|---|---|---|---|
| unsloth-UD-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.011657 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.042145 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.027578 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_S at 8.95 GiB | Shared experiment | 1 | 8.95 | 0.154054 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_XS at 8.17 GiB | Shared experiment | 1 | 8.17 | 0.226329 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_S at 11.29 GiB | Shared experiment | 1 | 11.29 | 0.059444 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_XXS at 9.73 GiB | Shared experiment | 1 | 9.73 | 0.112349 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ2_S at 10.38 GiB | Shared experiment | 1 | 10.38 | 0.1187 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ3_S at 12.89 GiB | Shared experiment | 1 | 12.89 | 0.044124 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ4_XS-IQ3_S at 13.45 GiB | Shared experiment | 1 | 13.45 | 0.033508 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.037988 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.025473 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.011143 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.015308 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.014556 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.061716 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ3_XS at 10.7 GiB | Shared experiment | 1 | 10.7 | 0.071421 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_XXS at 10 GiB | Shared experiment | 1 | 10 | 0.098482 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.055328 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiB | Shared experiment | 1 | 8.237 | 0.279307 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiB | Shared experiment | 1 | 9.238 | 0.183171 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiB | Shared experiment | 1 | 9.642 | 0.147895 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_S-3.23bpw at 10.277 GiB | Shared experiment | 1 | 10.28 | 0.104496 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiB | Shared experiment | 1 | 12.18 | 0.05204 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ2_S at 8.948 GiB | Shared experiment | 3 | 8.948 | 0.141824 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
- aee16ef6-dab6-4186-b9ed-1139db7e15d6: Superseded by q3-k-xl-ap
Fidelity vs file size, neutral web
Colour: packager.
Frontier chosen by hand: 8 runs, joined in file size (GiB) order. The dashed line connects recorded runs; it does not imply results between them.
How this chart is drawn
One point per successful run, no aggregation. Context series separate experiments, recorded comparison context, input references and environments; matching metadata does not establish experimental equivalence.
Schema versions contributing: quant-fidelity v1: 0 runs · quant-fidelity v2 (pinned): 26 runs. Versions join only where every charted field still records the same kind of value in the same unit and direction.
Displayed Y range: 0.008303 to 0.2403 · X range: 8.17 to 16.35
Schema 3dac891d-f938-47d2-84fc-304eb64adb94 · 25 eligible runs · 1 excluded · Live as of 2026-09-25T17:47:39.320Z
Source data and excluded runs
| Run | Experiment | Revision | file size (GiB) | KL divergence vs BF16 (lower is better) | Series | Context | Frontier |
|---|---|---|---|---|---|---|---|
| unsloth-UD-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.008695 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.027046 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.018594 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_S at 8.95 GiB | Shared experiment | 1 | 8.95 | 0.114309 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_XS at 8.17 GiB | Shared experiment | 1 | 8.17 | 0.172378 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_S at 11.29 GiB | Shared experiment | 1 | 11.29 | 0.043154 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_XXS at 9.73 GiB | Shared experiment | 1 | 9.73 | 0.082381 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ2_S at 10.38 GiB | Shared experiment | 1 | 10.38 | 0.085755 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ3_S at 12.89 GiB | Shared experiment | 1 | 12.89 | 0.030302 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ4_XS-IQ3_S at 13.45 GiB | Shared experiment | 1 | 13.45 | 0.023429 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.025002 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.018139 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.008303 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.010839 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.010443 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.040384 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ3_XS at 10.7 GiB | Shared experiment | 1 | 10.7 | 0.050795 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_XXS at 10 GiB | Shared experiment | 1 | 10 | 0.065797 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.037115 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiB | Shared experiment | 1 | 8.237 | 0.240308 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiB | Shared experiment | 1 | 9.238 | 0.151436 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiB | Shared experiment | 1 | 9.642 | 0.115114 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_S-3.23bpw at 10.277 GiB | Shared experiment | 1 | 10.28 | 0.077028 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiB | Shared experiment | 1 | 12.18 | 0.035723 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ2_S at 8.948 GiB | Shared experiment | 3 | 8.948 | 0.097686 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
- aee16ef6-dab6-4186-b9ed-1139db7e15d6: Superseded by q3-k-xl-ap
Top-1 agreement vs file size, held-out prose
Colour: packager.
Frontier chosen by hand: 8 runs, joined in file size (GiB) order. The dashed line connects recorded runs; it does not imply results between them.
How this chart is drawn
One point per successful run, no aggregation. Context series separate experiments, recorded comparison context, input references and environments; matching metadata does not establish experimental equivalence.
Schema versions contributing: quant-fidelity v1: 0 runs · quant-fidelity v2 (pinned): 26 runs. Versions join only where every charted field still records the same kind of value in the same unit and direction.
Displayed Y range: 0.779 to 0.9275 · X range: 8.17 to 16.35
Schema 3dac891d-f938-47d2-84fc-304eb64adb94 · 25 eligible runs · 1 excluded · Live as of 2026-09-25T17:47:39.320Z
Source data and excluded runs
| Run | Experiment | Revision | file size (GiB) | top-1 token agreement with BF16 (higher is better) | Series | Context | Frontier |
|---|---|---|---|---|---|---|---|
| unsloth-UD-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.92664 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.88776 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| unsloth-UD-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.90432 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_S at 8.95 GiB | Shared experiment | 1 | 8.95 | 0.81817 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ2_XS at 8.17 GiB | Shared experiment | 1 | 8.17 | 0.79084 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_S at 11.29 GiB | Shared experiment | 1 | 11.29 | 0.8729399999999999 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| GSQ-RCO-IQ3_XXS at 9.73 GiB | Shared experiment | 1 | 9.73 | 0.84379 | packager: ista-daslab | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ2_S at 10.38 GiB | Shared experiment | 1 | 10.38 | 0.83565 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ3_S at 12.89 GiB | Shared experiment | 1 | 12.89 | 0.8855500000000001 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AD-IQ4_XS-IQ3_S at 13.45 GiB | Shared experiment | 1 | 13.45 | 0.8978700000000001 | packager: atomicchat | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q3_K_XL at 12.24 GiB | Shared experiment | 1 | 12.24 | 0.89211 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ4_XS at 13.27 GiB | Shared experiment | 1 | 13.27 | 0.9042 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-Q4_K_XL at 16.35 GiB | Shared experiment | 1 | 16.35 | 0.92753 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.92069 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-Q4_K_M at 15.33 GiB | Shared experiment | 1 | 15.33 | 0.92235 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| unsloth-UD-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.8720300000000001 | packager: unsloth | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ3_XS at 10.7 GiB | Shared experiment | 1 | 10.7 | 0.86649 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_XXS at 10 GiB | Shared experiment | 1 | 10 | 0.8499 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| AP-IQ3_S at 11.21 GiB | Shared experiment | 1 | 11.21 | 0.87926 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
| ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiB | Shared experiment | 1 | 8.237 | 0.779 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiB | Shared experiment | 1 | 9.238 | 0.816 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiB | Shared experiment | 1 | 9.642 | 0.82789 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ3_S-3.23bpw at 10.277 GiB | Shared experiment | 1 | 10.28 | 0.84881 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiB | Shared experiment | 1 | 12.18 | 0.8855 | packager: byteshape | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | |
| AP-IQ2_S at 8.948 GiB | Shared experiment | 3 | 8.948 | 0.82612 | packager: agention | hardware="strix-halo-vulkan" · llamacpp_build="26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47" | Yes |
- aee16ef6-dab6-4186-b9ed-1139db7e15d6: Superseded by q3-k-xl-ap
Results
| Run | size gib (GiB) ↓ better | eff bpw (bits/weight) ↓ better | kld heldout ↓ better | kld wikitext ↓ better | kld mixedweb | kld heldout pm | kld mixedweb pm | kld wikitext pm | top1 heldout | top1 mixedweb | top1 wikitext | bytes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AD-IQ2_S at 10.38 GiB | 10.38 | 3.263 | 0.1187 | 0.106169 | 0.085755 | 0.000921 | 0.000868 | 0.001553 | 0.83565 | 0.85925 | 0.86676 | Missing |
| AD-IQ3_S at 12.89 GiB | 12.89 | 4.052 | 0.044124 | 0.044145 | 0.030302 | 0.000412 | 0.000315 | 0.001212 | 0.88555 | 0.90635 | 0.91105 | Missing |
| AD-IQ4_XS-IQ3_S at 13.45 GiB | 13.45 | 4.228 | 0.033508 | 0.038386 | 0.023429 | 0.000319 | 0.000236 | 0.001173 | 0.89787 | 0.91566 | 0.9189 | Missing |
| AP-IQ2_S at 8.948 GiB | 8.948 | 2.813 | 0.141824 | 0.116202 | 0.097686 | 0.001135 | 0.001003 | 0.001896 | 0.82612 | 0.84616 | 0.86282 | 9,607,980,000 |
| AP-IQ3_S at 11.21 GiB | 11.21 | 3.53 | 0.055328 | 0.049259 | 0.037115 | 0.000478 | 0.000422 | 0.001313 | 0.87926 | 0.90117 | 0.90971 | 12,040,900,000 |
| AP-IQ3_XS at 10.7 GiB | 10.7 | 3.36 | 0.071421 | 0.064237 | 0.050795 | 0.000667 | 0.000597 | 0.00132 | 0.86649 | 0.88503 | 0.89712 | 11,489,000,000 |
| AP-IQ3_XXS at 10 GiB | 10 | 3.14 | 0.098482 | 0.080097 | 0.065797 | 0.000821 | 0.000728 | 0.001554 | 0.8499 | 0.87204 | 0.88591 | 10,737,400,000 |
| AP-IQ4_XS at 13.27 GiB | 13.27 | 4.17 | 0.025473 | 0.024264 | 0.018139 | 0.000243 | 0.000229 | 0.000984 | 0.9042 | 0.92572 | 0.9328 | Missing |
| AP-Q3_K_XL at 12.24 GiB | 12.24 | 3.85 | 0.037988 | 0.03585 | 0.025002 | 0.000347 | 0.000283 | 0.001271 | 0.89211 | 0.91395 | 0.92138 | Missing |
| AP-Q4_K_M at 15.33 GiB | 15.33 | 4.82 | 0.014556 | 0.014985 | 0.010443 | 0.000166 | 0.00012 | 0.000897 | 0.92235 | 0.93969 | 0.94687 | Missing |
| AP-Q4_K_XL at 16.35 GiB | 16.35 | 5.14 | 0.011143 | 0.01181 | 0.008303 | 0.000108 | 0.000067 | 0.000712 | 0.92753 | 0.94238 | 0.95099 | Missing |
| GSQ-RCO-IQ2_S at 8.95 GiB | 8.95 | 2.813 | 0.154054 | 0.147016 | 0.114309 | 0.001254 | 0.001039 | 0.002052 | 0.81817 | 0.83297 | 0.84413 | Missing |
| GSQ-RCO-IQ2_XS at 8.17 GiB | 8.17 | 2.568 | 0.226329 | 0.203598 | 0.172378 | 0.001756 | 0.001441 | 0.002398 | 0.79084 | 0.80047 | 0.82276 | Missing |
| GSQ-RCO-IQ3_S at 11.29 GiB | 11.29 | 3.549 | 0.059444 | 0.066547 | 0.043154 | 0.000562 | 0.000438 | 0.001641 | 0.87294 | 0.89218 | 0.89573 | Missing |
| GSQ-RCO-IQ3_XXS at 9.73 GiB | 9.73 | 3.058 | 0.112349 | 0.106256 | 0.082381 | 0.000976 | 0.000809 | 0.001824 | 0.84379 | 0.85653 | 0.86813 | Missing |
| ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiB | 8.237 | 2.59 | 0.279307 | 0.274614 | 0.240308 | 0.002183 | 0.001928 | 0.002699 | 0.779 | 0.77195 | 0.79063 | 8,843,950,000 |
| ShapeLearn-IQ3_S-3.23bpw at 10.277 GiB | 10.277 | 3.23 | 0.104496 | 0.114002 | 0.077028 | 0.000984 | 0.00084 | 0.00182 | 0.84881 | 0.86442 | 0.86937 | 11,035,400,000 |
| ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiB | 9.642 | 3.03 | 0.147895 | 0.144414 | 0.115114 | 0.001351 | 0.001079 | 0.001935 | 0.82789 | 0.83501 | 0.84658 | 10,353,000,000 |
| ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiB | 9.238 | 2.9 | 0.183171 | 0.182559 | 0.151436 | 0.001636 | 0.001391 | 0.0022 | 0.816 | 0.81685 | 0.83051 | 9,919,480,000 |
| ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiB | 12.184 | 3.84 | 0.05204 | 0.053068 | 0.035723 | 0.000574 | 0.000419 | 0.001355 | 0.8855 | 0.90347 | 0.90894 | 13,083,100,000 |
| unsloth-UD-IQ3_S at 11.21 GiB | 11.21 | 3.53 | 0.061716 | 0.047003 | 0.040384 | 0.000542 | 0.000467 | 0.001188 | 0.87203 | 0.89891 | 0.91077 | 12,040,900,000 |
| unsloth-UD-IQ4_XS at 13.27 GiB | 13.27 | 4.17 | 0.027578 | 0.025212 | 0.018594 | 0.000279 | 0.000203 | 0.001056 | 0.90432 | 0.92398 | 0.93203 | Missing |
| unsloth-UD-Q3_K_XL at 12.24 GiB | 12.24 | 3.85 | 0.042145 | 0.033686 | 0.027046 | 0.000381 | 0.000297 | 0.001011 | 0.88776 | 0.91241 | 0.92183 | Missing |
| unsloth-UD-Q4_K_M at 15.33 GiB | 15.33 | 4.82 | 0.015308 | 0.013943 | 0.010839 | 0.000166 | 0.000109 | 0.000833 | 0.92069 | 0.93687 | 0.94798 | 16,464,400,000 |
| unsloth-UD-Q4_K_XL at 16.35 GiB | 16.35 | 5.14 | 0.011657 | 0.012169 | 0.008695 | 0.000119 | 0.000083 | 0.000912 | 0.92664 | 0.94194 | 0.95041 | 17,559,200,000 |
Finished runs and what they recorded.
AD-IQ2_S at 10.38 GiBsize gib 10.38
- size gib (GiB) ↓
- 10.38
- eff bpw (bits/weight) ↓
- 3.263
- kld heldout ↓
- 0.1187
- kld wikitext ↓
- 0.106169
- kld mixedweb
- 0.085755
- kld heldout pm
- 0.000921
- kld mixedweb pm
- 0.000868
- kld wikitext pm
- 0.001553
- top1 heldout
- 0.83565
- top1 mixedweb
- 0.85925
- top1 wikitext
- 0.86676
- bytes
- Missing
AD-IQ3_S at 12.89 GiBsize gib 12.89
- size gib (GiB) ↓
- 12.89
- eff bpw (bits/weight) ↓
- 4.052
- kld heldout ↓
- 0.044124
- kld wikitext ↓
- 0.044145
- kld mixedweb
- 0.030302
- kld heldout pm
- 0.000412
- kld mixedweb pm
- 0.000315
- kld wikitext pm
- 0.001212
- top1 heldout
- 0.88555
- top1 mixedweb
- 0.90635
- top1 wikitext
- 0.91105
- bytes
- Missing
AD-IQ4_XS-IQ3_S at 13.45 GiBsize gib 13.45
- size gib (GiB) ↓
- 13.45
- eff bpw (bits/weight) ↓
- 4.228
- kld heldout ↓
- 0.033508
- kld wikitext ↓
- 0.038386
- kld mixedweb
- 0.023429
- kld heldout pm
- 0.000319
- kld mixedweb pm
- 0.000236
- kld wikitext pm
- 0.001173
- top1 heldout
- 0.89787
- top1 mixedweb
- 0.91566
- top1 wikitext
- 0.9189
- bytes
- Missing
AP-IQ2_S at 8.948 GiBsize gib 8.948
- size gib (GiB) ↓
- 8.948
- eff bpw (bits/weight) ↓
- 2.813
- kld heldout ↓
- 0.141824
- kld wikitext ↓
- 0.116202
- kld mixedweb
- 0.097686
- kld heldout pm
- 0.001135
- kld mixedweb pm
- 0.001003
- kld wikitext pm
- 0.001896
- top1 heldout
- 0.82612
- top1 mixedweb
- 0.84616
- top1 wikitext
- 0.86282
- bytes
- 9,607,980,000
AP-IQ3_S at 11.21 GiBsize gib 11.21
- size gib (GiB) ↓
- 11.21
- eff bpw (bits/weight) ↓
- 3.53
- kld heldout ↓
- 0.055328
- kld wikitext ↓
- 0.049259
- kld mixedweb
- 0.037115
- kld heldout pm
- 0.000478
- kld mixedweb pm
- 0.000422
- kld wikitext pm
- 0.001313
- top1 heldout
- 0.87926
- top1 mixedweb
- 0.90117
- top1 wikitext
- 0.90971
- bytes
- 12,040,900,000
AP-IQ3_XS at 10.7 GiBsize gib 10.7
- size gib (GiB) ↓
- 10.7
- eff bpw (bits/weight) ↓
- 3.36
- kld heldout ↓
- 0.071421
- kld wikitext ↓
- 0.064237
- kld mixedweb
- 0.050795
- kld heldout pm
- 0.000667
- kld mixedweb pm
- 0.000597
- kld wikitext pm
- 0.00132
- top1 heldout
- 0.86649
- top1 mixedweb
- 0.88503
- top1 wikitext
- 0.89712
- bytes
- 11,489,000,000
AP-IQ3_XXS at 10 GiBsize gib 10
- size gib (GiB) ↓
- 10
- eff bpw (bits/weight) ↓
- 3.14
- kld heldout ↓
- 0.098482
- kld wikitext ↓
- 0.080097
- kld mixedweb
- 0.065797
- kld heldout pm
- 0.000821
- kld mixedweb pm
- 0.000728
- kld wikitext pm
- 0.001554
- top1 heldout
- 0.8499
- top1 mixedweb
- 0.87204
- top1 wikitext
- 0.88591
- bytes
- 10,737,400,000
AP-IQ4_XS at 13.27 GiBsize gib 13.27
- size gib (GiB) ↓
- 13.27
- eff bpw (bits/weight) ↓
- 4.17
- kld heldout ↓
- 0.025473
- kld wikitext ↓
- 0.024264
- kld mixedweb
- 0.018139
- kld heldout pm
- 0.000243
- kld mixedweb pm
- 0.000229
- kld wikitext pm
- 0.000984
- top1 heldout
- 0.9042
- top1 mixedweb
- 0.92572
- top1 wikitext
- 0.9328
- bytes
- Missing
AP-Q3_K_XL at 12.24 GiBsize gib 12.24
- size gib (GiB) ↓
- 12.24
- eff bpw (bits/weight) ↓
- 3.85
- kld heldout ↓
- 0.037988
- kld wikitext ↓
- 0.03585
- kld mixedweb
- 0.025002
- kld heldout pm
- 0.000347
- kld mixedweb pm
- 0.000283
- kld wikitext pm
- 0.001271
- top1 heldout
- 0.89211
- top1 mixedweb
- 0.91395
- top1 wikitext
- 0.92138
- bytes
- Missing
AP-Q4_K_M at 15.33 GiBsize gib 15.33
- size gib (GiB) ↓
- 15.33
- eff bpw (bits/weight) ↓
- 4.82
- kld heldout ↓
- 0.014556
- kld wikitext ↓
- 0.014985
- kld mixedweb
- 0.010443
- kld heldout pm
- 0.000166
- kld mixedweb pm
- 0.00012
- kld wikitext pm
- 0.000897
- top1 heldout
- 0.92235
- top1 mixedweb
- 0.93969
- top1 wikitext
- 0.94687
- bytes
- Missing
AP-Q4_K_XL at 16.35 GiBsize gib 16.35
- size gib (GiB) ↓
- 16.35
- eff bpw (bits/weight) ↓
- 5.14
- kld heldout ↓
- 0.011143
- kld wikitext ↓
- 0.01181
- kld mixedweb
- 0.008303
- kld heldout pm
- 0.000108
- kld mixedweb pm
- 0.000067
- kld wikitext pm
- 0.000712
- top1 heldout
- 0.92753
- top1 mixedweb
- 0.94238
- top1 wikitext
- 0.95099
- bytes
- Missing
GSQ-RCO-IQ2_S at 8.95 GiBsize gib 8.95
- size gib (GiB) ↓
- 8.95
- eff bpw (bits/weight) ↓
- 2.813
- kld heldout ↓
- 0.154054
- kld wikitext ↓
- 0.147016
- kld mixedweb
- 0.114309
- kld heldout pm
- 0.001254
- kld mixedweb pm
- 0.001039
- kld wikitext pm
- 0.002052
- top1 heldout
- 0.81817
- top1 mixedweb
- 0.83297
- top1 wikitext
- 0.84413
- bytes
- Missing
GSQ-RCO-IQ2_XS at 8.17 GiBsize gib 8.17
- size gib (GiB) ↓
- 8.17
- eff bpw (bits/weight) ↓
- 2.568
- kld heldout ↓
- 0.226329
- kld wikitext ↓
- 0.203598
- kld mixedweb
- 0.172378
- kld heldout pm
- 0.001756
- kld mixedweb pm
- 0.001441
- kld wikitext pm
- 0.002398
- top1 heldout
- 0.79084
- top1 mixedweb
- 0.80047
- top1 wikitext
- 0.82276
- bytes
- Missing
GSQ-RCO-IQ3_S at 11.29 GiBsize gib 11.29
- size gib (GiB) ↓
- 11.29
- eff bpw (bits/weight) ↓
- 3.549
- kld heldout ↓
- 0.059444
- kld wikitext ↓
- 0.066547
- kld mixedweb
- 0.043154
- kld heldout pm
- 0.000562
- kld mixedweb pm
- 0.000438
- kld wikitext pm
- 0.001641
- top1 heldout
- 0.87294
- top1 mixedweb
- 0.89218
- top1 wikitext
- 0.89573
- bytes
- Missing
GSQ-RCO-IQ3_XXS at 9.73 GiBsize gib 9.73
- size gib (GiB) ↓
- 9.73
- eff bpw (bits/weight) ↓
- 3.058
- kld heldout ↓
- 0.112349
- kld wikitext ↓
- 0.106256
- kld mixedweb
- 0.082381
- kld heldout pm
- 0.000976
- kld mixedweb pm
- 0.000809
- kld wikitext pm
- 0.001824
- top1 heldout
- 0.84379
- top1 mixedweb
- 0.85653
- top1 wikitext
- 0.86813
- bytes
- Missing
ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiBsize gib 8.237
- size gib (GiB) ↓
- 8.237
- eff bpw (bits/weight) ↓
- 2.59
- kld heldout ↓
- 0.279307
- kld wikitext ↓
- 0.274614
- kld mixedweb
- 0.240308
- kld heldout pm
- 0.002183
- kld mixedweb pm
- 0.001928
- kld wikitext pm
- 0.002699
- top1 heldout
- 0.779
- top1 mixedweb
- 0.77195
- top1 wikitext
- 0.79063
- bytes
- 8,843,950,000
ShapeLearn-IQ3_S-3.23bpw at 10.277 GiBsize gib 10.277
- size gib (GiB) ↓
- 10.277
- eff bpw (bits/weight) ↓
- 3.23
- kld heldout ↓
- 0.104496
- kld wikitext ↓
- 0.114002
- kld mixedweb
- 0.077028
- kld heldout pm
- 0.000984
- kld mixedweb pm
- 0.00084
- kld wikitext pm
- 0.00182
- top1 heldout
- 0.84881
- top1 mixedweb
- 0.86442
- top1 wikitext
- 0.86937
- bytes
- 11,035,400,000
ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiBsize gib 9.642
- size gib (GiB) ↓
- 9.642
- eff bpw (bits/weight) ↓
- 3.03
- kld heldout ↓
- 0.147895
- kld wikitext ↓
- 0.144414
- kld mixedweb
- 0.115114
- kld heldout pm
- 0.001351
- kld mixedweb pm
- 0.001079
- kld wikitext pm
- 0.001935
- top1 heldout
- 0.82789
- top1 mixedweb
- 0.83501
- top1 wikitext
- 0.84658
- bytes
- 10,353,000,000
ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiBsize gib 9.238
- size gib (GiB) ↓
- 9.238
- eff bpw (bits/weight) ↓
- 2.9
- kld heldout ↓
- 0.183171
- kld wikitext ↓
- 0.182559
- kld mixedweb
- 0.151436
- kld heldout pm
- 0.001636
- kld mixedweb pm
- 0.001391
- kld wikitext pm
- 0.0022
- top1 heldout
- 0.816
- top1 mixedweb
- 0.81685
- top1 wikitext
- 0.83051
- bytes
- 9,919,480,000
ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiBsize gib 12.184
- size gib (GiB) ↓
- 12.184
- eff bpw (bits/weight) ↓
- 3.84
- kld heldout ↓
- 0.05204
- kld wikitext ↓
- 0.053068
- kld mixedweb
- 0.035723
- kld heldout pm
- 0.000574
- kld mixedweb pm
- 0.000419
- kld wikitext pm
- 0.001355
- top1 heldout
- 0.8855
- top1 mixedweb
- 0.90347
- top1 wikitext
- 0.90894
- bytes
- 13,083,100,000
unsloth-UD-IQ3_S at 11.21 GiBsize gib 11.21
- size gib (GiB) ↓
- 11.21
- eff bpw (bits/weight) ↓
- 3.53
- kld heldout ↓
- 0.061716
- kld wikitext ↓
- 0.047003
- kld mixedweb
- 0.040384
- kld heldout pm
- 0.000542
- kld mixedweb pm
- 0.000467
- kld wikitext pm
- 0.001188
- top1 heldout
- 0.87203
- top1 mixedweb
- 0.89891
- top1 wikitext
- 0.91077
- bytes
- 12,040,900,000
unsloth-UD-IQ4_XS at 13.27 GiBsize gib 13.27
- size gib (GiB) ↓
- 13.27
- eff bpw (bits/weight) ↓
- 4.17
- kld heldout ↓
- 0.027578
- kld wikitext ↓
- 0.025212
- kld mixedweb
- 0.018594
- kld heldout pm
- 0.000279
- kld mixedweb pm
- 0.000203
- kld wikitext pm
- 0.001056
- top1 heldout
- 0.90432
- top1 mixedweb
- 0.92398
- top1 wikitext
- 0.93203
- bytes
- Missing
unsloth-UD-Q3_K_XL at 12.24 GiBsize gib 12.24
- size gib (GiB) ↓
- 12.24
- eff bpw (bits/weight) ↓
- 3.85
- kld heldout ↓
- 0.042145
- kld wikitext ↓
- 0.033686
- kld mixedweb
- 0.027046
- kld heldout pm
- 0.000381
- kld mixedweb pm
- 0.000297
- kld wikitext pm
- 0.001011
- top1 heldout
- 0.88776
- top1 mixedweb
- 0.91241
- top1 wikitext
- 0.92183
- bytes
- Missing
unsloth-UD-Q4_K_M at 15.33 GiBsize gib 15.33
- size gib (GiB) ↓
- 15.33
- eff bpw (bits/weight) ↓
- 4.82
- kld heldout ↓
- 0.015308
- kld wikitext ↓
- 0.013943
- kld mixedweb
- 0.010839
- kld heldout pm
- 0.000166
- kld mixedweb pm
- 0.000109
- kld wikitext pm
- 0.000833
- top1 heldout
- 0.92069
- top1 mixedweb
- 0.93687
- top1 wikitext
- 0.94798
- bytes
- 16,464,400,000
unsloth-UD-Q4_K_XL at 16.35 GiBsize gib 16.35
- size gib (GiB) ↓
- 16.35
- eff bpw (bits/weight) ↓
- 5.14
- kld heldout ↓
- 0.011657
- kld wikitext ↓
- 0.012169
- kld mixedweb
- 0.008695
- kld heldout pm
- 0.000119
- kld mixedweb pm
- 0.000083
- kld wikitext pm
- 0.000912
- top1 heldout
- 0.92664
- top1 mixedweb
- 0.94194
- top1 wikitext
- 0.95041
- bytes
- 17,559,200,000
The parameters of each run
| Run | arm | corpus | ctx | head variant | map lineage | packager | reference | tier | hardware (context) | llamacpp build (context) |
|---|---|---|---|---|---|---|---|---|---|---|
| unsloth-UD-Q4_K_XL at 16.35 GiB | as-shipped | both | 2,048 | base | unsloth-ud | unsloth | qwen38-27b-base-bf16 | unsloth-UD-Q4_K_XL | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| unsloth-UD-Q3_K_XL at 12.24 GiB | as-shipped | both | 2,048 | base | unsloth-ud | unsloth | qwen38-27b-base-bf16 | unsloth-UD-Q3_K_XL | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| unsloth-UD-IQ4_XS at 13.27 GiB | as-shipped | both | 2,048 | base | unsloth-ud | unsloth | qwen38-27b-base-bf16 | unsloth-UD-IQ4_XS | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| GSQ-RCO-IQ2_S at 8.95 GiB | as-shipped | both | 2,048 | base | ista-gsq-rco | ista-daslab | qwen38-27b-base-bf16 | Qwen3.8-27B-GSQ-RCO-IQ2_S-mtp | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| GSQ-RCO-IQ2_XS at 8.17 GiB | as-shipped | both | 2,048 | base | ista-gsq-rco | ista-daslab | qwen38-27b-base-bf16 | Qwen3.8-27B-GSQ-RCO-IQ2_XS-mtp | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| GSQ-RCO-IQ3_S at 11.29 GiB | as-shipped | both | 2,048 | base | ista-gsq-rco | ista-daslab | qwen38-27b-base-bf16 | Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| GSQ-RCO-IQ3_XXS at 9.73 GiB | as-shipped | both | 2,048 | base | ista-gsq-rco | ista-daslab | qwen38-27b-base-bf16 | Qwen3.8-27B-GSQ-RCO-IQ3_XXS-mtp | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AD-IQ2_S at 10.38 GiB | as-shipped | both | 2,048 | base | atomicchat-ad | atomicchat | qwen38-27b-base-bf16 | Qwen3.8-27B-AD-IQ2_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AD-IQ3_S at 12.89 GiB | as-shipped | both | 2,048 | base | atomicchat-ad | atomicchat | qwen38-27b-base-bf16 | Qwen3.8-27B-AD-IQ3_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AD-IQ4_XS-IQ3_S at 13.45 GiB | as-shipped | both | 2,048 | base | atomicchat-ad | atomicchat | qwen38-27b-base-bf16 | Qwen3.8-27B-AD-IQ4_XS-IQ3_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-Q3_K_XL at 12.24 GiB | as-shipped | both | 2,048 | base | unsloth-ud | agention | qwen38-27b-base-bf16 | AP-Q3_K_XL | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-IQ4_XS at 13.27 GiB | as-shipped | both | 2,048 | base | unsloth-ud | agention | qwen38-27b-base-bf16 | AP-IQ4_XS | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-Q4_K_XL at 16.35 GiB | as-shipped | both | 2,048 | base | unsloth-ud | agention | qwen38-27b-base-bf16 | AP-Q4_K_XL | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| unsloth-UD-Q4_K_M at 15.33 GiB | as-shipped | both | 2,048 | base | unsloth-ud | unsloth | qwen38-27b-base-bf16 | unsloth-UD-Q4_K_M | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-Q4_K_M at 15.33 GiB | as-shipped | both | 2,048 | base | unsloth-ud | agention | qwen38-27b-base-bf16 | AP-Q4_K_M | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| unsloth-UD-IQ3_S at 11.21 GiB | as-shipped | both | 2,048 | base | unsloth-ud | unsloth | qwen38-27b-base-bf16 | unsloth-UD-IQ3_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-IQ3_XS at 10.7 GiB | as-shipped | both | 2,048 | base | agention-precision | agention | qwen38-27b-base-bf16 | AP-IQ3_XS | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-IQ3_XXS at 10 GiB | as-shipped | both | 2,048 | base | agention-precision | agention | qwen38-27b-base-bf16 | AP-IQ3_XXS | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-IQ3_S at 11.21 GiB | as-shipped | both | 2,048 | base | unsloth-ud | agention | qwen38-27b-base-bf16 | AP-IQ3_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| ShapeLearn-IQ2_XXS-2.56bpw at 8.237 GiB | as-shipped | both | 2,048 | base | byteshape-shapelearn | byteshape | qwen38-27b-base-bf16 | Qwen3.8-27B-ShapeLearn-IQ2_XXS-2.56bpw | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| ShapeLearn-IQ3_XXS-2.88bpw at 9.238 GiB | as-shipped | both | 2,048 | base | byteshape-shapelearn | byteshape | qwen38-27b-base-bf16 | Qwen3.8-27B-ShapeLearn-IQ3_XXS-2.88bpw | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| ShapeLearn-IQ3_XS-3.01bpw at 9.642 GiB | as-shipped | both | 2,048 | base | byteshape-shapelearn | byteshape | qwen38-27b-base-bf16 | Qwen3.8-27B-ShapeLearn-IQ3_XS-3.01bpw | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| ShapeLearn-IQ3_S-3.23bpw at 10.277 GiB | as-shipped | both | 2,048 | base | byteshape-shapelearn | byteshape | qwen38-27b-base-bf16 | Qwen3.8-27B-ShapeLearn-IQ3_S-3.23bpw | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| ShapeLearn-IQ4_XS-3.84bpw at 12.184 GiB | as-shipped | both | 2,048 | base | byteshape-shapelearn | byteshape | qwen38-27b-base-bf16 | Qwen3.8-27B-ShapeLearn-IQ4_XS-3.84bpw | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
| AP-IQ2_S at 8.948 GiB | as-shipped | both | 2,048 | base | ista-gsq-rco | agention | qwen38-27b-base-bf16 | AP-IQ2_S | strix-halo-vulkan | 26bc85e42 plus local patches; llama-perplexity rebuilt 2026-09-18 22:47 |
1 run is not shown: unfinished or replaced
- AP-Q3_K_XL at 12.24 GiB succeededsuperseded
How this was measured
- Goal
Establish, under one stated protocol, how Agention Precision (AP-) GGUF tiers for Qwen3.8-27B compare with every other published quantization of the same model at matched file size. A comparison counts only when the two files are within about 1 percent on bytes.
- Success criteria
A tier wins its size slot when it beats the best rival file within 1 percent of its byte count on BOTH gating corpora, held-out and neutral web, by at least 3 sigma AND at least 10 percent relative. wikitext-2 is reported but does not gate. A win on one gating corpus only is recorded as a sub-bar result, not a ship claim.
Objective
One place for every file, ours and every rival, measured on one protocol so the frontier can be read directly. Roster: Agention Precision tiers, unsloth UD, IST-DASLab GSQ-RCO, ByteShape ShapeLearn, AtomicChat AD. Each row carries file bytes, effective bits per weight, mean KL divergence with its standard error on three corpora, and top-1 token agreement.
Method
Every file is measured against the same BF16 reference with llama-perplexity --kl-divergence at ctx 2048 over 60 chunks, on three corpora: held-out technical prose, a neutral FineWeb slice, and wikitext-2. Rival files are measured exactly as shipped, never rebuilt. Size comparisons are made against the nearest rival file by byte count, or against a log-interpolated vendor frontier when no file sits within 1 percent. Interpolated comparisons are labelled as such and never extrapolated beyond a vendor largest measured file.
Constraints
One 27B model fits on the GPU at a time, so all measurements are serial. Every file is measured against the same BF16 reference with llama-perplexity --kl-divergence over three corpora: held-out technical prose, a neutral FineWeb slice, and wikitext-2, at 60 chunks of 2048 tokens each. Two of the three corpora are reproducible by a reader; the held-out set is internal engineering documentation, so it is pinned by checksum rather than released, and any claim resting on it is flagged as such. Quantization method internals are withheld until the Qwen 4 release and are referenced through artifacts rather than described in the records.