Datasetpublic

phaedawg/qwen3.6-35b-a3b-distribution-fidelity-768x2048-v1

Qwen3.6-35B-A3B quantization analysis Mean KL divergence against on-disk size Scored under the distribution-fidelity laws, version 15. Read LAWS.md first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere. Each candidate directory holds its one-pager (report.md), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. reference/ carries the… See the full description on the dataset page: https://huggingface.co/datasets/phaedawg/qwen3.6-35b-a3b-distribution-fidelity-768x2048-v1.

sourceHugging Faceupdated 15d agoView on Hugging Face

0likes286downloads

Dataset Card

Qwen3.6-35B-A3B quantization analysis

Mean KL divergence against on-disk size

Scored under the distribution-fidelity laws, version 15. Read `LAWS.md` first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.

Each candidate directory holds its one-pager (report.md), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. reference/ carries the reusable teacher tensors and head, so a third party can score a new candidate without loading the reference checkpoint. checksums.txt covers every file here.

Candidates by paired routed-model fidelity

Candidate	Scheme	On disk	QxQ KLD	BxQ KLD	QxQ - BxQ	Natural route flips	Exact repeat
Qwen/Qwen3.6-35B-A3B-FP8	fp8_block	34.89 GiB	0.01730080	0.00757658	+0.00972422	35.9288%	certified
cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit	int4g32asym	23.25 GiB	0.02212258	0.01204155	+0.01008103	36.7551%	certified
QuantTrio/Qwen3.6-35B-A3B-AWQ	int4g128asym	23.71 GiB	0.02803013	0.01718293	+0.01084720	40.5507%	certified
palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4	int4g128sym	22.74 GiB	0.03866157	0.02633642	+0.01232514	43.8776%	certified
Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound	int4g128sym	20.02 GiB	0.04276821	0.02939575	+0.01337246	52.2533%	certified
unsloth/Qwen3.6-35B-A3B-NVFP4	nvfp4	24.67 GiB	0.04603616	0.03232773	+0.01370843	51.7594%	certified
nvidia/Qwen3.6-35B-A3B-NVFP4	nvfp4	21.82 GiB	0.04652691	0.03404794	+0.01247897	47.0071%	certified
unsloth/Qwen3.6-35B-A3B-NVFP4-Fast	nvfp4	22.02 GiB	0.05785633	0.04407521	+0.01378113	52.3272%	certified
RedHatAI/Qwen3.6-35B-A3B-NVFP4	nvfp4	23.32 GiB	0.06896001	0.05404429	+0.01491572	56.7272%	certified

QxQ fidelity against size

[image]

BxQ fidelity against size

[image]

The numbers behind the charts are in `kld-vs-size.json`, because a picture is not evidence.

Explicitly excluded checkpoints

These pinned revisions were deliberately not scored. They are listed so a known-bad checkpoint is never silently retried or mistaken for an omitted result.

—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.