Short answer (as of Oct 15, 2026): An uncensored model is one that says no less often. It is not smarter. It is just less careful. Most of the time it is a remix made by a hobbyist, not by the company that built the model.
What do uncensored and abliterated mean?
Most AI models are trained to say no to some questions. Think of how to build a bomb.
An uncensored model has had that training weakened. Abliterated is a nerdier word for one way to do it. Someone finds the part of the model that says no and switches it off.
Most of these are made by hobbyists. They take a model from a big company, change it, and post it under their own name. The big company did not make that version and does not stand behind it.
Is it smarter?
No. Taking out the no button does not add any brains. It can even make the model a bit worse at other things. It is the same model with fewer manners.
Why it matters for trust
A model that says yes to anything can help with bad stuff, and it will not warn you. If you share a computer with kids, or use AI for work, you may want a model with limits.
A remix also comes from a stranger, not the maker. That is a trust question too. Every model we list comes from the maker's own account.
Our labels
- Plays it safe: says no to touchy questions a lot. (0 models: none so far)
- Middle: answers most things and skips the really bad stuff. (21 models)
- Says yes to almost anything: almost no limits. It's not smarter, just less careful. (3 models: Mistral 7B v0.3, Mistral Nemo 12B, Hermes 3 8B)
- Not sure yet: the maker does not say, so we won't guess. (22 models)
Some models are sold as uncensored but still say no a lot. Those get a tag: Says it's uncensored, but isn't. None of the models we list get that tag right now.
How we pick a label
We only use what the maker wrote. That means the model card, a safety report, or the maker's own blog. Each model page links to the exact page we used.
Some makers say very little. Then we say "Not sure yet." We do not guess.
To be honest: we have not run our own tests yet. We plan to. We will ask each model the same set of questions and count how often it says no. Until then, these labels are the maker's word, not ours.
The label sits next to the trust score. It does not change it.
Common questions
Is an uncensored AI model smarter?
No. It just says no less often. Removing refusals does not add knowledge, and can make it a little worse at other tasks.
What does abliterated mean?
It is one way to make an uncensored model. Someone finds the part of the model that refuses and switches it off. It is usually done by hobbyists, not the maker.
Did Trusty Models test these models?
Not yet. Our labels come from what each maker wrote. We plan to run our own tests.