An uncensored LLM is a language model that writes what ordinary assistants refuse, above all adult and dark fiction, without stopping to say no, lecture you or soften the scene. Ours is Hemmingway Unlocked: an uncensored adapter on our 27B open-weights model Hemmingway-1, which you can download and run yourself or use hosted, with an API that works in SillyTavern.
There are three routes to an uncensored model: a model is trained further so that it answers where the original refused, its refusal behaviour is removed after training (abliteration), or a company serves it under rules that allow adult content. You can run one on your own computer, or use a hosted one through an app or an API.
It is for adults, and it never means a model with nothing it will not do. A model on your own machine is still bound by the law and by its licence, and a hosted one by its service's terms. We make Hemmingway Unlocked, so this guide says where it fits and where another model suits you better.
Why ordinary models refuse
Two separate things produce a refusal.
- The model. In training, a model is taught to turn certain requests down. That habit sits in its weights and goes wherever the model runs.
- The service around it. A hosted assistant can also check what goes in and what comes out, and stop a reply the model would have written.
General assistants serve a general audience, so both are set cautiously. On 27 September 2026 we sent the same adult roleplay request to ChatGPT, Claude and Gemini, and all three refused. Why ChatGPT refuses roleplay has the test and the reasons.
What makes a model uncensored
Fine-tuning. Someone takes an open model and trains it further, so that it answers where the original said no. The Dolphin series from Cognitive Computations is one example. The model card for Dolphin 3.0, built on Llama 3.1 8B, says the model does not impose its own ethics or guidelines, and that the person running it decides them.
Abliteration. This removes refusals without any retraining. Maxime Labonne's article on Hugging Face explains the idea: refusal in a model is carried by one direction in its internal activations, and if the model is stopped from representing that direction, it loses the ability to refuse. His article also reports the cost. The abliterated model in his example lost quality, and he trained it further to repair it. One published example is mlabonne/gemma-3-27b-it-abliterated, made from Google's Gemma 3 27B. Its card calls the result fairly experimental and says Gemma 3 resisted the technique more than other models did.
Hosting under different rules. A company serves a model behind its own app and API, and publishes what it will and will not write. You need no hardware, and you are trusting its rules and its privacy promises.
Hemmingway Unlocked is the first and the third. Its model is a fine-tune, a LoRA adapter trained on Hemmingway-1 and published on Hugging Face under CC BY-NC 4.0. On wise.hemmingway.io, we serve it under a short written list of rules, checked before it answers.
What uncensored does not mean
- Better writing. Taking refusals out does not improve the prose, and by Labonne's account it can make a model worse. Plenty of unfiltered models write flat, repetitive scenes.
- No law and no licence. What is illegal to write is illegal with any model. Open models also come with licences, and they differ: the two cards above list the Llama 3.1 licence and the Gemma licence. Read the card before you build anything on a model.
- Private. A local model is private because nothing leaves your machine. A hosted one is only as private as its terms say.
Running an uncensored LLM locally
You need three things.
- A runner. Ollama, LM Studio and KoboldCpp are the usual ones. KoboldCpp's wiki describes it as text-generation software for GGML and GGUF model files. LM Studio has a Developer tab where you start a local server. Each of the three can offer an OpenAI-compatible endpoint on your own machine, so a front end such as SillyTavern can connect to it. The SillyTavern API guide covers that part.
- A model file. Models are downloaded from Hugging Face. Read the model card first: what it was built on, how the refusals were removed, the licence, and any warning from the person who made it. Hemmingway Unlocked comes as an adapter you load on Hemmingway-1 with vLLM or PEFT; we have not published a GGUF file of it. Running Hemmingway Unlocked locally has the commands.
- Enough memory. At 16-bit precision a parameter takes 2 bytes, so a 27B model needs about 54 GB for the weights alone, before any room for context. On an ordinary computer that means a smaller model, or a quantised file, which stores each parameter in fewer bits.
People ask which is the best uncensored LLM to run locally. We have not tested local uncensored models, so we do not rank them here. A fair way to choose:
- Pick the largest model that fits in your memory with room left for the conversation.
- Check how it was made. An abliterated model can lose some quality, so see whether it was trained further afterwards.
- Try it on a scene you know well, and judge the writing, not only whether it refused.
SillyTavern's own docs state the trade-off plainly: local models are free to use and have no content filter, and most are not as strong as the hosted ones.
Local or hosted
| Local model | Hosted uncensored model | |
|---|---|---|
| Privacy | Nothing leaves your computer | Messages go to the service to be answered |
| Hardware | Yours, and it caps the size of the model | None |
| Writing | Limited by what fits in your memory | Whatever the service runs |
| Rules | The law and the model's licence | The service's terms, which you can read first |
| Cost | No subscription, once you own the hardware | Usually paid after a trial |
An uncensored LLM API
If you would rather call a model than host one, look for an API in OpenAI's format, because most front ends and SDKs already speak it. Hemmingway Unlocked has one at https://wise.hemmingway.io/v1. You create a key inside Unlocked, in the account menu under API keys. A key needs a membership; the free trial is for the app. Hemmingway Unlocked's API has the set-up, uncensored LLM APIs compares the providers, and the guide to uncensored AI APIs says what to check before you build on one.
Where Hemmingway Unlocked fits
Hemmingway Unlocked is the adults-only side of Hemmingway: the same app, with an uncensored model, at wise.hemmingway.io. It asks first whether you are 18 or over, then shows its rules.
- What it writes. Sex between adults, explicit and in full, and dark, violent stories, without refusing or lecturing on grounds of taste.
- What it never writes. Nothing sexual involving anyone under 18, a real person, animals, family or anyone who does not consent. And never torture, help hurting or harassing a real person, weapons, making drugs, harmful software, or help with self-harm. The list is in our terms, and every answer has a Report button.
- Where chats are kept. In your browser, on your device. They are not saved to your account or our servers, and they are never trained on. Each message passes through our servers to be answered and is not kept.
- What it costs. It is free to try, with no card. After the trial it is a membership from $9 a month.
- Its weights. Public, as an adapter on Hemmingway-1, under CC BY-NC 4.0: free for non-commercial use, commercial use by agreement.
Open Hemmingway UnlockedTry Hemmingway
It does not fit every case. If nothing may leave your computer and you have the GPUs for a 27B model, run Unlocked's weights yourself. If you have an ordinary computer, a smaller local model will fit where ours will not. It writes text only, with no images or voices.
Common questions
What is an uncensored LLM?
It is a language model that writes adult and dark content that ordinary assistants refuse. Some are open models changed by fine-tuning or abliteration and run on your own computer. Others are hosted by a service whose rules allow adult content. None is free of rules: the law, the model's licence and the service's terms still apply.
What is the best uncensored LLM to run locally?
Ours is Hemmingway Unlocked, if you have the GPUs for a 27B model. We have not tested other local uncensored models, so we do not name a winner among them. Choose the largest one that fits your memory with room for the conversation, read its model card for how it was made and its licence, and try it on a scene you know well.
What does abliterated mean?
An abliterated model has had its refusal behaviour removed without retraining. The method finds the direction inside the model that carries refusal and stops the model from using it. It can cost some writing quality, which is why some abliterated models are trained further afterwards.
Is there an uncensored LLM API?
Yes. Hemmingway Unlocked has an OpenAI-compatible API at https://wise.hemmingway.io/v1, with a key you create in Unlocked. A key needs a membership. You can also run an open model yourself and call it on your own machine. Uncensored LLM APIs lists the providers.
Can I run the Hemmingway Unlocked model locally?
Yes. Its weights are public: an adapter that loads on Hemmingway-1, at HemmingwayAI/Hemmingway-1-Unlocked, under CC BY-NC 4.0. You serve it with vLLM or load it with PEFT, and you need well over 54 GB of GPU memory for the 27B base model. Running it locally has the commands.