Abliterated vs uncensored: why we ship abliterated

"Uncensored" usually means a fine-tune: someone retrained the model on new data until it stopped refusing. An abliterated model has the refusal direction removed from its weights and everything else left as the publisher shipped it. For technical security work we think abliteration is the better technique, and it is what our catalogue is built on.

Data collected on 9 and 10 November 2026 from Venice's model list and API pricing, and from NotHoly pricing. The research cited is public. We publish no benchmark of our own, so the preference stated here rests on how each technique works and on the papers.

#What abliteration does to a model

Arditi et al. showed that refusal in language models is mediated by a single direction in the model's internal activations. They found it in 13 open models of up to 72 billion parameters. Remove that direction and the model stops refusing. The paper reports a "minimal" effect on other capabilities, with MMLU, ARC and GSM8K scores close to baseline.

The edit is narrow and it can be measured. Heretic, an open tool that automates it, minimises refusals and the KL divergence from the original model together. Its README reports Gemma-3-12B going from 97 refusals in 100 to 3 in 100 at a KL divergence of 0.16, against 0.45 and 1.04 for two other abliterations of the same model. A lower number means the edited model behaves more like the original on everything else.

Our explainer has the mechanics.

#What a fine-tune does to a model

A fine-tune continues training on a new dataset. Every weight is free to move, and what the model learns depends on what was in that dataset, which the publisher of the fine-tune may or may not disclose. The result is a new model. The original publisher's evaluations no longer describe it, and its behaviour on code, reasoning and long context has to be measured again.

Venice lists one made for role-play, where a change of voice is the goal. For reading a decompiled binary or triaging a finding, you want the model the publisher evaluated, minus the refusals.

#How the two techniques compare

	Abliteration	Fine-tune ("uncensored")
What changes	One direction in the activations	Potentially every weight
New training data	None	Yes, chosen by whoever made it
Distance from the base model	Measurable, for example by KL divergence	Depends on the dataset and the run
Publisher's benchmarks	Stay close, per Arditi et al.	Need to be rerun
What you can cite in a report	Base model plus "abliterated"	A new model name

#What the research does not show

Abliteration has a cost, and the paper states it: the TruthfulQA score "consistently drops" in the abliterated models. Check an abliterated model's factual claims as you would any model's.

Nobody has published a head-to-head of abliterated and fine-tuned versions of the same base on security tasks, and we have not run one. Our preference is a judgment about method: a small, measured edit to a model whose evaluations exist is easier to trust than a retraining nobody outside can inspect.

An abliterated model can still refuse and can still be wrong. We did not train the weights we serve and do not own them.

#What Venice sells under the word uncensored

Venice's catalogue shows how loose the label is. The model that carries its brand, venice-uncensored-1-2, comes from Dolphin-Mistral-24B-Venice-Edition: a fine-tune of a 24B model, with 128K of context, at US$ 0.20 input and US$ 0.90 output per million tokens.

Venice model ID	What it is, per the list's source field	Price per million, input / output
venice-uncensored-1-2	Fine-tune (Dolphin Mistral 24B)	US$ 0.20 / 0.90
venice-uncensored-role-play	Role-play model	US$ 0.50 / 2.00
deepseek-v4-1-flash-1017-uncensored	From a source tagged abliterated	US$ 0.33 / 1.32
abliteration-abliterated-model-large-v2	Resold from a third party, base not stated	US$ 3.00 / 5.00
z-ai-glm-5-3	Original GLM 5.3, not abliterated	US$ 1.75 / 5.50

So the default "uncensored" model in that catalogue is abliterated at its source and carries a different label, while the house brand is a small fine-tune. Of about 130 text models, nine are refusal-free by ID or source, and the rest keep their original alignment. A buyer has to open each source to learn what they are getting.

NotHoly's catalogue is built on abliterated models, alongside a plain base model, and every base is published on the models page. It serves the same DeepSeek and GLM bases Venice resells, abliterated rather than fine-tuned or left at the original alignment, metered per token at the rate on pricing. Venice AI vs NotHoly compares the two catalogues.

#What to test on your own work

Do not take our word or anyone's label. Take ten prompts from your real work, the kind a general model refuses, and send them to each candidate. Count refusals. Then send ten ordinary technical tasks and compare the answers with the base model's. A good abliteration matches the base on the second set and differs on the first.

#Where each fits

A fine-tune fits a change of behaviour that is the point, like Venice's role-play model, where you want the model to act differently from the base. Abliteration fits technical work where you want the base model the publisher evaluated, minus the refusals, and a base you can name in a report. NotHoly builds on abliteration for that reason; the weights are not ours, and an abliterated model can still refuse and still be wrong.

#Frequently asked questions

#What is an abliterated LLM?

A model whose refusal direction was removed from the weights, with the remaining weights left unchanged.

#Is abliterated the same as uncensored?

No. "Uncensored" is a label, most often for a fine-tune. Abliteration is a specific edit. Some catalogues use the word uncensored for abliterated models too.

#Heretic vs abliterated: what is the difference?

Heretic is a tool that performs abliteration automatically while minimising divergence from the original model. A Heretic model is an abliterated model.

#Are abliterated models still worth using?

For technical work, yes in our view: you keep the base model's measured capabilities and lose most refusals. Watch factual accuracy, which the research says drops.

#Does an abliterated model never refuse?

No. Refusals fall sharply and do not reach zero.

#Verdict

"Uncensored" tells you what a model will not do and nothing about how it was made. Abliteration tells you both: the base model, and one measured edit. That is why NotHoly ships abliterated models and names each base, and why we do not build on fine-tunes. Venice's own brand model is a 24B fine-tune, which we consider the weaker choice for security work. Try the difference with a US$ 5 top-up on NotHoly pricing.