If you want an uncensored LLM API for adult fiction and roleplay in text, in OpenAI's format, with open weights behind it and a key that works in SillyTavern, the one we make is Hemmingway Unlocked. If you need images or audio from the same key, hundreds of models behind one account, or a model on your own hardware, one of the other providers below will suit you better.
This page is the list: who offers one, what each one's own pages say about its rules and what it keeps, and whether it works in SillyTavern. We make Hemmingway Unlocked. What we say about the others comes from their own sites, read on 27 October 2026. For what to check before you build, see uncensored AI API. For a pick for SillyTavern, see the best API for SillyTavern.
What an uncensored LLM provider is
It runs a model that writes adult and dark content without refusing, and gives you a key to call it from your own code or a front end. Some serve their own uncensored model, with one set of rules. Others are hosts or marketplaces, where what you may write depends on the model you pick and who serves it. Or you serve an open model yourself. Every provider still has rules: "uncensored" means fewer refusals, not none.
The providers side by side
| Provider | What it is | OpenAI-compatible | What it keeps, per its own pages | In SillyTavern |
|---|---|---|---|---|
| Hemmingway Unlocked (ours) | One uncensored model for adults, open weights | Yes | Messages answered, not kept, never trained on | Custom OpenAI-compatible source |
| Venice | Private AI with text, image, video and audio APIs | Yes | Nothing stored on normal requests; on some models an outside provider sees the prompt | Not in SillyTavern's docs |
| OpenRouter | Hundreds of models from many providers, one API | Yes | Nothing stored unless you opt in; each provider has its own policy | Built-in source |
| Featherless | Hosting for open-weight models | Yes | No logs of what is sent through its API | Its docs offer a managed SillyTavern |
| NanoGPT | Many closed and open models, one account | Yes | API logging off by default; each model's provider still gets the request | Built-in source |
| ArliAI | A subscription for open and tuned models | Yes | No logs of requests or generations | Its quick start covers it |
| Your own hardware | vLLM, Ollama or KoboldCpp | Yes | Nothing leaves your machine | KoboldCpp and Ollama are local sources |
Where Hemmingway Unlocked fits
Hemmingway Unlocked is the adults-only side of Hemmingway. Its API serves the same uncensored model as its app.
- Address:
https://wise.hemmingway.io/v1, in OpenAI's Chat Completions format. - Key: made inside Unlocked, in the account menu under API keys. A key needs a membership, from $9 a month; the free trial is for the app. An Unlocked key always reaches Unlocked's model, whatever model name you send.
- SillyTavern: Chat Completion, Custom (OpenAI-compatible), the address and your key. The set-up guide goes step by step.
- What it writes: sex between adults, explicit and in full; dark, violent fiction; roleplay; ordinary chat.
- What it never writes: anything sexual involving anyone under 18 (or an adult playing someone under 18), a real person, animals, family or anyone who does not want it or cannot say yes; and never torture, help hurting a real person, stalking, doxxing, harassing or scamming, weapons, explosives, making drugs, harmful software, or help with suicide or self-harm. The full list is in our terms. Your app inherits it.
- What is kept: each message passes through our servers to be answered and is not kept, and nothing is used for training.
- Open weights: the model is an adapter (LoRA) on Hemmingway-1, our 27B open-weights model built on Qwen3.8-27B. It is on Hugging Face under CC BY-NC 4.0: free for non-commercial use, commercial use by agreement. You can move to your own hardware later.
- Text only. No images, no audio.
Open Hemmingway UnlockedTry Hemmingway
The other providers
Venice
Venice's API docs call it a privacy-first, uncensored AI API with no content filtering, across text, image, video and audio. It works with OpenAI's SDK at the base URL https://api.venice.ai/api/v1.
- Rules. Its terms of service, last updated 8 June 2026, rule out illegal use, abusing or exploiting children, including AI-generated child sexual abuse material, and using anyone's name, voice, image or likeness without their consent.
- What it keeps. Its privacy docs say Venice's proxy stores no prompt or reply. On "anonymized" models an outside company runs the model: Venice hides your identity from it, and its privacy page says to assume that company stores your prompt. On "private" models nothing is retained.
- SillyTavern. SillyTavern's docs do not name it.
OpenRouter
OpenRouter's docs describe one API for hundreds of models. It is a marketplace, so what is allowed depends on the model and the host. Its terms say what you send must follow the terms of the model or provider you use, and SillyTavern's docs describe it as having no enforced moderation unless the model's maker requires it.
- What it keeps. Its data collection page says it stores no prompts or responses unless you opt in, and keeps metadata such as token counts. Each provider has its own logging and retention policy, and you can turn off providers that may train on your data.
- Format. OpenAI's SDK works with the base URL
https://openrouter.ai/api/v1. - SillyTavern. A built-in Chat Completion source.
Featherless
Featherless calls itself a serverless inference platform for open-weight models, with a large catalogue you call through one key. Its site lets you filter that catalogue for uncensored models.
- Rules. Its terms say the models on it may produce content not suitable for minors, that you must follow the terms of each model you use, and that you may not use it to break the law or other people's rights.
- What it keeps. Its privacy docs say it does not log chats, prompts or completions sent through its API.
- Format. OpenAI-compatible, at
https://api.featherless.ai/v1. - SillyTavern. Its docs offer a managed SillyTavern with Featherless already connected, on its higher paid plans.
NanoGPT
NanoGPT puts ChatGPT, Claude, Gemini and hundreds more models behind one account, with an OpenAI-compatible API. By its own account it runs no models itself: each request goes to the provider of the model you pick.
- Rules. Its terms, last updated 24 October 2026, are for adults 18 and over. They forbid sexual content involving anyone under 18 or anyone who appears to be, sexually explicit content of an identifiable person without consent, and stalking, doxxing, threats and fraud.
- What it keeps. Its privacy page says chats in its web app stay in your browser by default, and that API logging is off by default. The provider of each model still receives the request, and what is kept varies by provider.
- SillyTavern. SillyTavern's docs list it among their Chat Completion sources, and NanoGPT has its own guide.
ArliAI
ArliAI's site calls it unrestricted AI inference: a monthly subscription for open models and tuned versions of them, with text and image generation.
- Rules. Its terms are for adults 18 and over, and rule out unlawful, fraudulent or harmful use, naming hate, harassment, violence, political manipulation, spam and malware.
- What it keeps. Its site says it keeps no logs of requests or generations, and that chats in its own app are stored on your device.
- Format. The standard OpenAI format, by its site.
- SillyTavern. Its roleplay quick start covers SillyTavern, by Chat Completion or Text Completion.
Your own hardware: vLLM, Ollama, KoboldCpp
Each serves an open-weight model on your own machine, and each one's docs give an OpenAI-compatible endpoint. vLLM's server supports OpenAI's Chat Completions API. Ollama's is at http://localhost:11434/v1/. KoboldCpp, a single program with nothing to install, serves one at http://localhost:5001/v1.
- Rules. Only the model's licence and the law. The model's training still decides what it refuses.
- What it keeps. Nothing leaves your machine.
- SillyTavern. Its docs list KoboldCpp and Ollama as local APIs, free and with no content filter, and name vLLM as well.
The cost is hardware. Unlocked's own weights need about 54 GB for the weights alone at 16-bit.
Where Unlocked does not fit
- You need images, video or audio from one key. Unlocked writes text only. Venice's docs cover all four, and NanoGPT and ArliAI offer image generation.
- You want to switch between many models. Unlocked is one model. OpenRouter and NanoGPT put hundreds behind one key.
- You want a character library, avatars or an app-store app. Unlocked has none of these. You describe the character in your first message.
- You want the API without paying. A key needs a membership. The free trial is for the app.
- You want to use the weights commercially. CC BY-NC 4.0 covers non-commercial use. Commercial use is by agreement with us.
How to choose
- Adult fiction and roleplay in text, open weights, one model with its rules in writing: Hemmingway Unlocked.
- Text, images, video and audio from one API: Venice.
- Hundreds of models behind one key: OpenRouter or NanoGPT. Read the rules of the model you pick.
- Open-weight models you choose, hosted for you: Featherless or ArliAI.
- Nothing may leave your machine: vLLM, Ollama or KoboldCpp on your own hardware.
Common questions
Which providers offer an uncensored LLM API?
Hemmingway Unlocked, which is ours. Venice and ArliAI, which call themselves uncensored or unrestricted. OpenRouter, NanoGPT and Featherless, where it depends on the model you pick. And any open model you serve yourself. Each still has rules, so read them before you build.
Is there an uncensored LLM API that is OpenAI-compatible?
Yes, every option on this page. Hemmingway Unlocked's is at https://wise.hemmingway.io/v1: in OpenAI's SDK you change the base URL and the key, and the rest of the call stays as it was.
Which uncensored LLM APIs work in SillyTavern?
SillyTavern's docs list OpenRouter and NanoGPT as built-in sources, and KoboldCpp, Ollama and vLLM as local ones. Featherless and ArliAI have SillyTavern guides of their own. Hemmingway Unlocked goes in as a custom OpenAI-compatible source.
Can I run Hemmingway Unlocked's model myself?
Yes. Its weights are public, as an adapter on Hemmingway-1 under CC BY-NC 4.0, and need a strong GPU set-up. Run Hemmingway Unlocked locally shows how.