phaedawg/distribution-fidelity-index
Local Inference Lab: distribution-fidelity index Every published one-pager, with a pointer to the full artifact that backs it. Each artifact repo carries the token suite and the reusable reference tensors, so a third party can score a new candidate without loading the reference checkpoint. Read LAWS.md, currently version 15, before comparing anything here. Numbers are comparable only within a single artifact's suite, geometry, and runtime identity. Each one-pager cites the laws… See the full description on the dataset page: https://huggingface.co/datasets/phaedawg/distribution-fidelity-index.
Local Inference Lab: distribution-fidelity index
Every published one-pager, with a pointer to the full artifact that backs it. Each artifact repo carries the token suite and the reusable reference tensors, so a third party can score a new candidate without loading the reference checkpoint.
Read `LAWS.md`, currently version 15, before comparing anything here. Numbers are comparable only within a single artifact's suite, geometry, and runtime identity. Each one-pager cites the laws version it was scored under, which is not always the newest.
Qwen3.6-27B
Artifact: `phaedawg/qwen3.6-27b-distribution-fidelity-768x2048-v1`
- Qwen FP8 — QxQ KLD 0.01631336
- nvidia NVFP4 — QxQ KLD 0.04045750
- cyankiwi AWQ-INT4 — QxQ KLD 0.04152462
- Lorbus int4-AutoRound — QxQ KLD 0.05100551
- unsloth NVFP4 — QxQ KLD 0.05139168
- mmangkad NVFP4 — QxQ KLD 0.05664272
- QuantTrio AWQ — QxQ KLD 0.06023338
- sudo-0x2a NVFP4-GPTQ — QxQ KLD 0.07933551
- rdtand PrismaSCOUT-Blackwell-NVFP4-BF16-vllm — QxQ KLD 0.08449741
- groxaxo GPTQ-Pro-4bit — QxQ KLD 0.12576406
Qwen3.6-35B-A3B
Artifact: `phaedawg/qwen3.6-35b-a3b-distribution-fidelity-768x2048-v1`
- Qwen FP8 — QxQ KLD 0.01730080, BxQ 0.00757658, QxQ-BxQ +0.00972422, exact-repeat certified
- cyankiwi AWQ-4bit — QxQ KLD 0.02212258, BxQ 0.01204155, QxQ-BxQ +0.01008103, exact-repeat certified
- QuantTrio AWQ — QxQ KLD 0.02803013, BxQ 0.01718293, QxQ-BxQ +0.01084720, exact-repeat certified
- palmfuture GPTQ-Int4 — QxQ KLD 0.03866157, BxQ 0.02633642, QxQ-BxQ +0.01232514, exact-repeat certified
- Intel int4-mixed-AutoRound — QxQ KLD 0.04276821, BxQ 0.02939575, QxQ-BxQ +0.01337246, exact-repeat certified
- unsloth NVFP4 — QxQ KLD 0.04603616, BxQ 0.03232773, QxQ-BxQ +0.01370843, exact-repeat certified
- nvidia NVFP4 — QxQ KLD 0.04652691, BxQ 0.03404794, QxQ-BxQ +0.01247897, exact-repeat certified
- unsloth NVFP4-Fast — QxQ KLD 0.05785633, BxQ 0.04407521, QxQ-BxQ +0.01378113, exact-repeat certified
- RedHatAI NVFP4 — QxQ KLD 0.06896001, BxQ 0.05404429, QxQ-BxQ +0.01491572, exact-repeat certified
Explicitly excluded pinned checkpoints:
ykarout/Qwen3.6-35B-A3B-NVFP4@5093c2a1d4d263007b563753438caf75dadaec48— Its forward pass emits NaN hidden states: every one of the 63569920 logits in the first scored window is non-finite, with the hidden states themselves already NaN before the head. No mean can be computed, so there is nothing to score rather than a poor score to report. Not a defect of this pipeline; the other NVFP4 candidates of this family score at the same tensor-parallel size.
Qwen3.8-27B
Artifact: `phaedawg/qwen3.8-27b-distribution-fidelity-768x2048-v1`
- Qwen FP8 — QxQ KLD 0.01097467
- cyankiwi AWQ-INT4 — QxQ KLD 0.03116102
- unsloth NVFP4 — QxQ KLD 0.04053080
- dbirks W4A16-AutoRound — QxQ KLD 0.04285275
- RadixArk NVFP4-BF16-LMHead — QxQ KLD 0.05156551
- philbert440 W4A16-AWQ — QxQ KLD 0.05220833
- RedHatAI INT4 — QxQ KLD 0.06290331
- RadixArk NVFP4 — QxQ KLD 0.06405373
- gittensor-model-hub NVFP4-RTX5090 — QxQ KLD 0.08890054
- Inferact NVFP4 — QxQ KLD 0.13893424
gemma-4-26B-A4B-it
Artifact: `phaedawg/gemma-4-26b-a4b-it-distribution-fidelity-768x2048-v1`
- RedHatAI FP8-dynamic — QxQ KLD 0.69415039, BxQ 0.49783177, QxQ-BxQ +0.19631862, exact-repeat certified
- cyankiwi AWQ-4bit — QxQ KLD 1.02626820, BxQ 0.84088786, QxQ-BxQ +0.18538034, exact-repeat certified
- unsloth NVFP4 — QxQ KLD 1.13846019, BxQ 0.95586495, QxQ-BxQ +0.18259525, exact-repeat certified
- dhruvil237 W4A16 — QxQ KLD 1.15926068, BxQ 1.01252351, QxQ-BxQ +0.14673717, exact-repeat certified
- Intel int4-AutoRound — QxQ KLD 1.17816288, BxQ 0.90308200, QxQ-BxQ +0.27508088, exact-repeat certified
- lcu0312 AWQ-4bit — QxQ KLD 1.23587221, BxQ 1.08009636, QxQ-BxQ +0.15577585, exact-repeat certified
- RedHatAI NVFP4 — QxQ KLD 1.77968754, BxQ 1.54324011, QxQ-BxQ +0.23644744, exact-repeat certified
- Neural-ICE Gemma-4-26B-A4B-it-NVFP4 — QxQ KLD 1.82281421, BxQ 1.58512109, QxQ-BxQ +0.23769312, exact-repeat certified
- bg-digitalservices Gemma-4-26B-A4B-it-NVFP4 — QxQ KLD 1.82307292, BxQ 1.59361302, QxQ-BxQ +0.22945990, exact-repeat certified
Explicitly excluded pinned checkpoints:
cklaus/gemma-4-26B-A4B-it-NVFP4@54218eb0fd33da17632f0d91be841d09a7724a6c— Scored on vLLM fdc0b57e while every other candidate scored on c798225a, so its report binds a reference capture the family no longer publishes (Law 12) and it ranks in a comparability group of one. Withdrawn rather than rescored: no measurement of this checkpoint is in question, only the provenance it carries.
gemma-4-31B-it-qat-q4_0-unquantized
Artifact: `phaedawg/gemma-4-31b-it-qat-q4_0-unquantized-distribution-fidelity-768x2048-v1`
- prithivMLmods gemma-4-31B-it-qat-FP8 — QxQ KLD 0.14337459
- cyankiwi gemma-4-31B-it-qat-AWQ-INT4 — QxQ KLD 0.22346921
- unsloth gemma-4-31B-it-qat-w4a16 — QxQ KLD 0.27404863
- google gemma-4-31B-it-qat-w4a16-ct — QxQ KLD 0.27404863
- yasu-oh gemma-4-31B-it-qat-NVFP4 — QxQ KLD 0.41421886
- melcheikh gemma-4-31B-it-qat-NVFP4-Blackwell — QxQ KLD 0.48811344
- melcheikh gemma-4-31B-it-qat-NVFP4-mse-Blackwell — QxQ KLD 0.48811344
