Abliterated vs uncensored
Both kinds of model answer where a stock model would refuse. An abliterated model has had the refusal cut out of it. An uncensored model was trained not to refuse in the first place.
Abliteration
A language model that refuses does it in a surprisingly simple way. Compare what happens inside the model on requests it refuses and requests it answers, and one direction in its internal activity lights up for the refusals. Abliteration finds that direction and removes it from the model's weights. Nothing is retrained. The model keeps what it knew and loses the reflex to say no.
The appeal is that it is quick, cheap, and can be done to any open model within days of release. The risk is a blunt edit: remove too much and the model gets slightly worse at everything.
Heretic
Heretic is a tool that automates abliteration. Instead of a person choosing where and how hard to cut, it tries many settings and keeps the one that stops the most refusals while changing the model's ordinary answers the least. The result is usually a lighter edit than one made by hand.
Uncensored fine-tunes
An uncensored model gets there by training. The model is trained further on conversations that contain no refusals, so it learns to answer. This takes more time and compute, and it can change the model's voice, for better or worse. Done well, it produces a model that is not only willing but good at the kind of writing it was trained on.
Which is better
- For staying close to the original model, abliterated. You get the stock model's abilities with the refusals gone.
- For roleplay and fiction, a good uncensored fine-tune often writes better, because it was trained on that material.
- In day to day chat, most people cannot tell them apart. Try both on the thing you actually want to do.
Which models here are which
| Model | How the refusals were removed |
|---|---|
| Kimi K3.1 Pro | Abliterated |
| DeepSeek V4.1 Flash 1017 | Abliterated |
| DeepSeek Flash Lite | Abliterated with Heretic |
| MiMo V2.6 Flash | Abliterated |
| GLM 5.3 and GLM 5.3 Flash | Uncensored build |
| Qwen3.8 27B | Uncensored build |
Hosted or downloaded
Abliterated models are popular as downloads to run with Ollama or llama.cpp. That works if you have the GPU for it. The larger models on this page need far more memory than a home machine has, which is the reason to use them hosted.
Keep reading
- Kimi K3.1 Pro, uncensored→
- GLM 5.3, uncensored→
- DeepSeek V4.1 Flash, uncensored→
- Qwen3.8 27B, uncensored→
- The uncensored AI API→
- Set up SillyTavern→
- Set up opencode→
- Set up Claude Code→
- Cost calculator→
- Uncensored AI roleplay→
Updated 12 November 2026. Prices are read from the live price list.