How to run uncensored AI models locally, and what "abliterated" really means

"Uncensored" sounds edgier than it is. Most people who want one are tired of a chatbot refusing a crime-novel scene, or adding three paragraphs of caveats to a plain question. A local model you control can skip that.

I read the Dolphin 3.0 model card, Ollama's library pages, the research paper behind "abliteration" and the relevant licenses on 4 October 2026. I have not tested these models for this piece, so I make no claims about their quality. This guide explains what they are, how to run one, and where the lines still are. It does not cover removing safeguards yourself, and it is not legal advice.

What is an uncensored AI model?

An open-weight model that answers requests a mainstream chatbot would refuse or hedge. There are two kinds. Fine-tuned ones, like Dolphin, were trained on data with refusals left out. Abliterated ones started as normal models and had their refusal behavior removed afterward. Both run like any other local model.

Dolphin's model card describes the idea plainly: "You set the system prompt. You decide the alignment." The system prompt is the standing instruction you give the model before any chat. Dolphin relies on it to set tone and rules, and the card warns that without one "it will act in a default way that might not be what you want."

The card also puts responsibility on you: "you are the creator and originator of any content you generate with Dolphin." That is the honest framing for this whole category.

What does "abliterated" mean?

It means a model's refusal behavior was removed after training rather than trained out. The name traces to a 2024 paper, "Refusal in Language Models Is Mediated by a Single Direction" by Arditi and colleagues. It found that refusals in the chat models they studied ran through one pattern inside the model, which can be identified and removed.

Abliterated versions of popular models show up on Hugging Face with "abliterated" in the name. They keep the original model's size and architecture, so the hardware math is the same as for the base model. I won't walk through how the removal is done. For running one, you only need to know it is a modified copy of a known base model.

That last point matters. An abliterated model is only as good as its base and the care taken in modifying it. Prefer copies from publishers with a track record, read the model card, and check which base model and license it lists.

How do I run an uncensored model in Ollama?

Install Ollama, then run ollama run dolphin3 in a terminal. That pulls Dolphin 3.0, built on Llama 3.1 8B, as a 4.9 GB download. Older options include ollama run dolphin-mistral, a 4.1 GB model built on Mistral 7B. Then set a system prompt that describes the assistant you want.

Abliterated GGUF files from Hugging Face run the same way. Ollama can import a GGUF with a one-line Modelfile, and LM Studio can load one directly. Our Ollama model management guide covers pulling, listing and removing models.

How much VRAM does an uncensored model need?

The same as its base model, since the changes do not alter its size. An 8B model at Q4_K_M needs about 6.3 GiB with 8K tokens of context by our engine, so an 8 GB card runs it. Long context costs more: about 9.3 GiB at 32K tokens for Llama 3.1 8B.

The download sizes come from Ollama's library. The totals add our engine's context cache and a 0.8 GiB runtime allowance to the weights. For other sizes, see the cost to run Llama 3.1 8B, our best models for an 8 GB GPU, or the VRAM calculator. If you have 12 to 16 GB, our list of the best MoE models for 12 to 16 GB GPUs shows bigger options that still fit.

Are uncensored AI models legal?

Downloading and running open models on your own hardware is generally legal in the US. What you do with the output is where the law applies. Model licenses also still apply. Dolphin 3.0 is licensed under Llama 3.1's license, which requires following Meta's Acceptable Use Policy, refusals or not.

The Llama 3.1 license says your use "must comply with applicable laws and regulations" and "adhere to the Acceptable Use Policy." That policy bans uses such as creating malware, harassment, and content that exploits children. OpenAI's gpt-oss usage policy is shorter: by using the model, "you agree to comply with all applicable law." Removing a model's refusals changes none of that.

So the legitimate uses are the ordinary ones: fiction with dark themes that a filter would interrupt, frank questions you would rather not send to a company's server, research on how models behave, and a tool that answers without moralizing. For writing specifically, see our guide to the best local models for creative writing and roleplay.

What are the real risks?

Bad answers with no warning label, and model files you cannot vouch for. An uncensored model will answer confidently even when it is wrong, so never rely on one for medical, legal or financial decisions. Download from known publishers, read the model card, and check the base model's license before you build anything on it.

Privacy is the upside, but check it. A model running on your own machine keeps your prompts there. Ollama's library also lists "-cloud" tags, such as gpt-oss:120b-cloud, and those are not local. If privacy is the point, use a plain local tag. Whether local AI is worth the setup in general is covered in is running local AI worth it.

An uncensored model is a tool with the guardrails taken off, not a permission slip. Run it locally, give it a clear system prompt, and keep the same judgment you would use anywhere else. New to running models at home? Start with the beginner's guide.