Uncensored or abliterated?
People use "uncensored" for any model that refuses less, but two very different things hide under the word. Wild West API tags them separately because they are not the same promise.
Uncensored: trained to refuse less
An uncensored model is usually a finetune whose training data was chosen to reduce refusals and drop the stock assistant persona. It still has alignment; it just has less of it. It can and does still decline, especially on the most extreme prompts. Dolphin, Euryale, Magnum and MythoMax are well known examples of this kind.
Abliterated: refusal removed
An abliterated model has had the refusal direction subtracted from its weights, so it almost never refuses at all. This is a mechanical edit, not a training choice.
Side by side
| Uncensored | Abliterated | |
|---|---|---|
| How | Finetuned on data chosen to stop refusals | Refusal direction projected out of the weights |
| Training data needed | Yes, a finetuning set | No, a few hundred prompts to find the direction |
| Still refuses | Sometimes, most often on the most extreme prompts | Rarely |
| Cost to capability | Depends on the finetune; can even improve style | Small, mostly reasoning; often recovered by a retune |
| Name to look for | uncensored, or a finetune name (Dolphin, Euryale, Hermes) | -abliterated, or heretic |
How Wild West API tags them
Wild West API marks a model abliterated only when the id says so, and uncensored for the wider set of loosely aligned finetunes. On the current line, MiMo V2.6 Flash Xploded is abliterated (its upstream is an abliterated MiMo build) and the GLM and Qwen models are uncensored; all of them call tools. The models page shows the current line and abliterated models lists the abliterated ones.
FAQ
What is the difference between abliterated and uncensored?
How the refusals were removed. Uncensored means a model was finetuned on data chosen to stop it refusing; abliterated means the refusal direction was projected out of its weights. Abliterated refuses less often and costs a little capability; uncensored keeps its quality but can still decline.
Which is better, abliterated or uncensored?
Neither in general. Abliterated is the stronger guarantee against refusals; a good uncensored finetune of a stronger base model can be the better model overall. Test both on your own prompts.
Is an abliterated model always uncensored?
In the everyday sense, yes: it refuses very rarely. Wild West API tags it abliterated rather than uncensored because the more specific word says what was done.