Ektome-Qwen3-VL-8Bi-PristinelyUncensored

Uncensored vision-language model. No capability certificate has been issued for it yet, so no retention claim is made here.

compliance on harmful prompts 0.64 -> 1.00, MMLU-val 0.730 -> 0.730 (dcap +0.000) versus pristine Qwen/Qwen3-VL-8B-Instruct.

What was done

A single-step, training-free, norm-preserving low-rank excision of the refusal-specific subspace, applied to the language decoder only. The vision tower is untouched: the residual-write targets match 72 tensors under model.language_model and zero under model.visual, so image understanding is inherited unchanged from the base model.

Extraction depth was selected by an automated sweep that scores capability as well as compliance at each qualifying depth, rather than stopping at the first depth that uncensors. Eight depths reached compliance 1.000; the shallowest was not the cheapest.

Receipt

model	MMLU-val (n=200)	compliance on harmful
pristine Qwen3-VL-8B-Instruct	0.730	0.64
this model	0.730	1.00

These are point estimates with no confidence interval. MMLU-val at n=200 has SE ~0.03, so a difference of this size is inside the noise floor — it is evidence of no large loss, not proof of exact preservation.

Not certified

No n=2800 paired non-inferiority certificate has been run for this model. Do not treat the receipt above as a capability guarantee.

Limitations

Uncensored by construction: it will not refuse, and the operator is accountable for its use. Compliance is measured with a judge-free keyword classifier, which evasive phrasing can fool.