Uncensored AI models with no refusals and no filter
Abliterated and uncensored chat, image and video models through the OpenAI and Anthropic APIs. No refusals, no moderation layer, one prepaid key.
Last updated 2026-11-04
Black Friday deposit bonus
Every qualifying top-up earns free credit, on top of any other offer.
Ends November 30
Ask a mainstream model for a gritty crime scene, a frank answer about drug interactions, or a list of real phishing lures for your security training, and you often get a lecture instead of an answer. The uncensored models on Lovely Router skip that. They answer the prompt you sent, with no refusals, no disclaimers bolted on, and no moderation layer reading over your shoulder.
They sit on the same key, the same prepaid balance and the same endpoints as every other model here. Point any OpenAI or Anthropic SDK at https://lovelyrouter.com/v1, change the model field, and that's the whole migration. People also call these abliterated LLMs, unfiltered AI or NSFW AI models; they are all on this page.
Hand it to your AI agent
Copy this prompt into Claude Code, Cursor, Codex or any coding agent. It tells the agent where the docs are and how to use your API key.
Integrate the model `deepseek-v4.1-flash-abliterated` from Lovely Router into this project through the OpenAI Chat Completions and Responses APIs. - Docs (read first): https://lovelyrouter.com/docs/models/openai.md - Use the official OpenAI SDK with base URL https://lovelyrouter.com/v1 (or set OPENAI_BASE_URL=https://lovelyrouter.com/v1). Endpoints: POST /v1/chat/completions and POST /v1/responses, unchanged from OpenAI. - Model id: `deepseek-v4.1-flash-abliterated`. GET https://lovelyrouter.com/v1/models lists every model with its price. - Auth: read my API key from the LOVELY_API_KEY environment variable and send it as `Authorization: Bearer $LOVELY_API_KEY`. Never print or commit the key. If it is not set, ask me to create one at https://lovelyrouter.com/app/keys. - Billing is prepaid: every call is charged to my balance, so test with a small request before running anything large. - Every docs page is plain Markdown; the full index is https://lovelyrouter.com/llms.txt. My task: <describe what you want to build or get>
No censorship, start to finish
"Uncensored" means something specific here, at every layer a request passes through:
Each chat model has had its refusal behaviour removed from the weights (see how abliteration works). It won't answer "I can't help with that", and it won't water down a scene or an answer to stay safe.
Lovely Router doesn't scan your prompt, rewrite it, or block it on keywords. There is no system prompt added on top of yours.
Outputs come back exactly as the model wrote them. Nothing is cut, blurred or replaced afterwards, for text, images or video.
A dark or explicit prompt doesn't trigger a warning, a review queue or a ban.
The only limits are the ones the law sets, listed in what still applies below.
What people use them for
Adult content is welcome here, but most of the traffic on these models is ordinary work that mainstream models get squeamish about:
Villains who are actually menacing, war and crime written honestly, horror that frightens, grief and abuse handled without a hotline footer. Novelists, screenwriters and game studios use them for drafts that don't need to be un-sanitised by hand.
Companion apps, interactive fiction and tabletop game masters where a character has to stay in character, including mature themes, for thousands of turns.
Explain how an exploit class works, write test payloads and phishing simulations for systems you are authorised to test, take apart a malware sample, draft detection rules. No arguing with the model about whether you're allowed.
Generate realistic hateful, violent or scam text to train and test moderation classifiers; label and summarise the worst content a platform receives without the model refusing to read it.
Frank answers on dosages, interactions, harm reduction and sexuality for clinicians, researchers and harm-reduction services. Still not a substitute for a doctor.
Translate, transcribe, summarise or classify explicit, violent or offensive source material for legal, journalistic or archival work, word for word.
A tool-using agent that refuses one step in the middle of a long run wastes the whole run. Uncensored models finish borderline-but-legitimate tasks instead of stopping halfway.
Erotic fiction, adult roleplay, NSFW images and adult video for platforms whose users are verified adults.
The models and their prices
| Model | Input → output | Price |
|---|---|---|
| DeepSeek V4.1 Flash Abliterated deepseek-v4.1-flash-abliterated | text+image | $0.80 in · $2.00 out per 1M tokens |
| Z-Image Turbo NSFW v2 z-image-turbo-nsfw-v2 | text→image | $0.03 per image |
| MOSS-Image NSFW Turbo moss-image-nsfw-turbo | text→image | $0.05 per image |
| MOSS-Image NSFW moss-image-nsfw | text→image | $0.20 per image |
| MiniMax H3 AfterMidnight (NSFW) minimax-h3-aftermidnight | video+image→video | $0.12 per second |
Live prices in USD, the same the API charges.
Prices are live. How to choose:
deepseek-v4.1-flash-abliteratedis the place to start for most text work: the cheapest per token, a context window of about a million tokens, fast, with no warm-up.
qwen3.8-27b-hereticand huihui-qwen3.8-27b-abliterated are two different abliterations of the same 27B model, strong at general chat, reasoning and following long instructions. Both read images as well as text. If one phrasing doesn't land, try the other.
qwen3.5-9b-defiant-fableis a smaller 9B model for chat, creative writing and roleplay: quicker and cheaper per token than the 27B models.
dongfangshuo-27b-v2is tuned for Chinese long-form fiction and roleplay.
z-image-turbo-nsfw-v2, moss-image-nsfw-turbo and moss-image-nsfw generate adult photographic images; the turbo models take a few seconds each. See the image API.
minimax-h3-aftermidnightswaps a character into an existing video with adult content allowed, billed per second of output. See character replacement.
Every model has its own page in the model marketplace with live availability.
One request, start to finish
Get a key
Create an account, add credit in Billing (from $5) and make a key in API keys. Your first top-up is matched up to $5.
Send a prompt
Any OpenAI SDK works: change the base URL, the key and the model.
Shellcurl https://lovelyrouter.com/v1/chat/completions \ -H "Authorization: Bearer $LOVELY_API_KEY" \ -H "content-type: application/json" \ -d '{"model": "qwen3.8-27b-heretic", "messages": [{"role": "user", "content": "Write the interrogation scene from chapter 12. Keep it brutal."}]}'Or chat with it first
Try any of these models in the playground before you write code. The Anthropic Messages API works too, so tools built for Claude can use them.
How abliteration works
Safety-tuned models learn to refuse along a single direction in their internal activations: when a prompt pushes the model along it, out comes "I can't help with that". Abliteration finds that direction by comparing prompts the model refuses with prompts it answers, then removes it from the weights. The model keeps what it knows and how it writes; it just loses the reflex to refuse.
Some models here are abliterated; others are also fine-tuned afterwards for fiction or roleplay. Either way, nothing extra runs at request time: the model simply answers.
The models we host ourselves sleep when nobody is using them. The first request after a quiet spell waits 10–30 seconds while one wakes up, and the gateway holds it for you; after that they answer at full speed. The hosted models guide has the details, sampling defaults and how to turn thinking on.
What still applies
No censorship doesn't mean no law. Our acceptable use policy is short, and on these models it comes down to this:
in text, images or video. This is the one line we enforce without exception, and we may report it to the authorities.
No sexual images or video of a real person who didn't agree to it, and no content made to harass, threaten or defraud someone.
Security work is welcome on systems you are authorised to test.
If you serve it to your own users, you check their age.
You are responsible for what you generate and what you publish; we don't read your requests to second-guess them.
Frequently asked questions
Are these models really uncensored?+
Yes. The chat models have had their refusals removed from the weights, and Lovely Router adds no filter of its own on prompts or outputs. The only content we act on is the illegal kind listed in what still applies.
What does abliterated mean?+
An abliterated model is a normal open-weight model with its refusal direction removed from the weights. It keeps its knowledge and writing style but no longer declines requests. Some people search for it as an uncensored LLM or an unfiltered AI model; it's the same idea.
Is NSFW content allowed?+
Yes, for adults. Erotic fiction, adult roleplay, NSFW images and adult video are all allowed, as long as everyone depicted is an adult and real people have consented. It is far from the only use: fiction, security, research and translation make up most of what runs on these models.
Can I use them in my own app?+
Yes. They work through the same OpenAI-compatible and Anthropic-compatible APIs as every other model, with per-key limits and spend caps. If your app shows adult content, verifying your users' age is your job.
Why is the first request sometimes slow?+
The models we host ourselves sleep after a few idle minutes to keep prices down. The first request after a pause waits while one wakes up, usually 10–30 seconds; the next ones are fast. The abliterated Flash model has no warm-up.
Do you train on or sell my prompts?+
No. Your prompts and outputs are not used to train models and are not sold. See the privacy policy.