What is an abliterated model?
An abliterated model is an open-weight language model whose refusal behavior was reduced by editing its weights. Research on refusal found that it is mediated by a single direction in the model's activations; abliteration removes that direction from the weights. No prompt trick and no retraining are involved.
What abliteration changes
--- base model (open weights)
+++ NotHoly model
Removed: refusal direction in the weights
Added: answers the request as written
what the base model knows, nothing new
its mistakes: check the output
your prompt, with nothing added to it
your authorization, your contract, our terms
--- base model (open weights)
+++ NotHoly model
Removed: refusal direction in the weights
Added: answers the request as written
what the base model knows, nothing new
its mistakes: check the output
your prompt, with nothing added to it
your authorization, your contract, our terms
# Source: Arditi et al., Refusal in Language Models Is Mediated by a Single Direction, 2024.
It answers instead of refusing.
Abliteration finds the direction in the weights that triggers refusals and removes it. Nothing is added to your prompt, so there is no jailbreak to break.
It knows what the base model knows.
No new knowledge and no benchmark of ours. It can still be wrong, so check its work against the evidence.
It refuses less, and sometimes still refuses.
Some refusals survive abliteration. When one shows up, make the scope explicit and ask again.
Authorization stays with you.
It will discuss the dangerous case. Whether you may test a system is up to your contract, your program's rules and our terms.
How it works
Researchers compare the model's internal activations on requests it refuses and requests it answers. The difference points along one direction. Projecting that direction out of the weights weakens refusals while leaving the rest of the model in place.
What it changes
The model answers requests that a stock assistant declines, including questions security work depends on, such as how a bypass works or what a binary does.
What it does not change
It adds no knowledge and fixes no mistakes. An abliterated model can be wrong in the same ways as the original, and it can still refuse some requests.
Abliteration and jailbreaks
A jailbreak is a prompt that tries to talk a model out of refusing and stops working when the model is updated. Abliteration is a change to the weights, made once.
Where NotHoly stands
NotHoly serves abliterated versions of open-weight models and names the base model of each. NotHoly does not train the weights. Use is limited to authorized work.
Source: Arditi et al., Refusal in Language Models Is Mediated by a Single Direction, 2024.
Updated 2026-11-06
Ask anything your scope allows.
Create an account, add credits from $10 and ask the first question in the chat or the terminal.
- 1/3
Create the account
Individual or company. You can switch to a company later without losing balance or keys.
IndividualCompany - 2/3
Add credits
By card, from $10. Pay per token, no subscription.
$10.00minimum deposit - 3/3
Ask
In the chat, or install the CLI and run notholy login in your terminal.
$ notholy login