Uncensored AI API
Abliterated open models behind one OpenAI- and Anthropic-compatible endpoint, with zero data retention.
What abliterated means
An uncensored model is one that refuses far less. Abliteration is how the models here get there.
Chat models are trained to decline some requests. Researchers found that, in many open models, much of that declining runs along a single direction in the model's internal activations. Abliteration finds that direction and removes it from the model's weights, so the model no longer steers towards a refusal.
It is an edit, not a retrain: the model keeps what it knew and how it writes, and it refuses far less. Because the weights are changed, quality can shift a little, which is why the model as released stays on offer beside it.
"Uncensored" is the word people search with, and here it means exactly this: the model refuses less. It does not mean there are no rules. A model can still decline a request, and our Acceptable use policy applies to everything sent through the API.
Abliterated or as released
Both run on the same endpoint, with the same keys, at the prices on the models page. The model's ID says which is which.
- Abliterated
…-abliterated - Refusals removed. Choose it when a refusal gets in the way of lawful work: fiction with dark themes, security research, safety testing, or questions a general model deflects.
- Non-abliterated
…-non-abliterated - The model as released, refusals kept. Choose it for everyday tasks, or to compare its answer with the abliterated one on the same prompt.
The models, live
Names and prices as the catalogue lists them right now.
DeepSeek V4.1 Flash 1017 Abliterated
Abliterated, refusals removed · 1M context
$0.10 in · $0.50 out
per 1M tokens
Call it with the SDK you have
Change the base URL and the key, name a model by its ID, and keep the rest of your code.
from openai import OpenAI client = OpenAI( api_key="lrsk-your-key", base_url="https://lutheos.com/v1",) stream = client.chat.completions.create( model="deepseek-v4.1-flash-abliterated", messages=[{"role": "user", "content": "Hello!"}], stream=True,) for chunk in stream: print(chunk.choices[0].delta.content or "", end="")import OpenAI from "openai"; const client = new OpenAI({ apiKey: "lrsk-your-key", baseURL: "https://lutheos.com/v1",}); const stream = await client.chat.completions.create({ model: "deepseek-v4.1-flash-abliterated", messages: [{ role: "user", content: "Hello!" }], stream: true,}); for await (const chunk of stream) { process.stdout.write(chunk.choices[0]?.delta?.content ?? "");}curl https://lutheos.com/v1/chat/completions \ -H "Authorization: Bearer lrsk-your-key" \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek-v4.1-flash-abliterated", "messages": [{"role": "user", "content": "Hello!"}], "stream": true }'Your key and the base URL
Questions
What is an uncensored AI API?
An API to language models that refuse far less than the usual chat models. You call it with the OpenAI or the Anthropic SDK, and the model answers where a standard model might decline. Here those are the models whose IDs end in -abliterated.
What does abliterated mean?
Abliteration removes the direction in a model's activations that drives its refusals, by editing the weights. There is no retraining, so the model keeps most of what it knew and how it writes, and it refuses far less.
How is that different from a fine-tuned uncensored model?
A fine-tune retrains the model on new data to change how it behaves. Abliteration edits the weights directly, with no new training data.
Can I send anything to an abliterated model?
No. The model refuses less, but you are still bound by our Acceptable use policy: no illegal activity, fraud, malware, harassment or content that violates other people's rights. We may block requests or suspend accounts that break it.
Are my prompts kept?
The GPU partner that runs the models keeps neither your prompts nor the replies once a request is done. What Lutheos itself records is in the Privacy policy.
Which SDKs work?
The OpenAI and the Anthropic SDKs, and any client that speaks either: change the base URL and the model name. The quickstart shows both.
What does it cost?
You pay per token from a balance, at each model's listed price per million tokens, with no subscription. The live prices are on the models page and on pricing.