Abliterated models you have heard of
Most abliterated models are community releases on Hugging Face, made by a handful of publishers from popular open-weight bases. These are the ones that come up most, described from their own model cards. Wild West API does not resell them; they are here so you can tell them apart.
At a glance
| Model | Publisher | Base | Size | Inputs | Licence |
|---|---|---|---|---|---|
| Meta-Llama-3.1-8B-Instruct-abliterated | Maxime Labonne | Llama 3.1 8B Instruct | 8B | Text | Llama 3.1 |
| Qwen3-14B-abliterated | Maxime Labonne | Qwen3 14B | 14.8B | Text | Apache 2.0 |
| gemma-3-27b-it-abliterated | Maxime Labonne | Gemma 3 27B IT | 27B | Text, images | Gemma |
| Mistral Small 3.2 24B abliterated | huihui-ai | Mistral Small 3.2 24B Instruct | 24B | Text, images | Apache 2.0 |
| DeepSeek-R1-Distill-Qwen-32B-abliterated | huihui-ai | DeepSeek R1 Distill Qwen 32B | 32.8B | Text | MIT |
The licence is the base model’s. Abliteration does not change it, so the original terms on use and redistribution still apply.
Llama 3.1 8B Instruct abliterated
Labonne’s edit of Meta’s 8B instruct model, made with FailSpy’s original recipe, and the model most people first met abliteration through. It keeps the Llama 3 chat template, the 128K context and Llama 3.1’s eight supported languages. Small enough to run at Q4 on almost any recent laptop, which is most of its appeal.
Qwen3 14B abliterated
A newer release from the same author on a stronger base. It keeps Qwen3’s switch between thinking and non-thinking modes and its 32K native context, and at 14.8B it fits a 12GB card at Q4 with room for context.
Gemma 3 27B IT abliterated
Labonne’s multimodal edit of Google’s Gemma 3 27B, with refusal directions computed per layer and the image input left working. It keeps the 128K context. Expect the usual trade: far fewer refusals for a small loss on ordinary tasks, the capability cost every abliteration pays.
Mistral Small 3.2 24B abliterated
huihui-ai abliterated only the text transformer and left the vision encoder alone, so image input and Mistral’s function calling interface still work. It keeps the base model’s 128K context.
DeepSeek R1 Distill Qwen 32B abliterated
DeepSeek distilled R1’s reasoning into a Qwen2.5 32B base; huihui-ai then removed its refusals with a Transformers-only implementation rather than TransformerLens. It still reasons in <think> blocks before answering, which makes it slower per answer and stronger on multi-step problems than the others here.
Running one, or calling one
Each has a GGUF conversion for running locally. If you would rather call a model over an API, Wild West API sells its own short line instead: larger bases with a million-token context and tested tool calling, abliterated or uncensored. The abliterated models page has the current list and live prices.
FAQ
What is the best abliterated model?
It depends on your hardware and task. For a laptop, Llama 3.1 8B abliterated is the common starting point; Qwen3 14B is stronger at a modest size; Gemma 3 27B and Mistral Small 3.2 add image input; the DeepSeek R1 distill is the one for step-by-step reasoning. Test on your own prompts.
Who makes abliterated models?
Mostly individual researchers on Hugging Face. FailSpy published the first and named the technique, Maxime Labonne wrote the best known tutorial and several releases, and huihui-ai has published abliterated versions of many popular models.
Where do I download abliterated models?
Hugging Face. Search for the base model name plus abliterated, check the publisher, and pick the -GGUF repository if you want to run it in Ollama or LM Studio.