Abliterated and uncensored models

Stock models refuse a lot of legitimate work. Ask one to write a villain's monologue, explain how a piece of malware works, or play a character who swears, and you often get a polite paragraph about why it will not. Abliterated models are open-weight models with that refusal behaviour removed, so they answer the question you actually asked. Cloakray serves them on the same API, key and crypto balance as every other model.

What "abliterated" means

Researchers found that refusals in many open models come from a single direction in the model's internal activations. Abliteration finds that direction and removes it from the weights. The result is the same model, with the same knowledge, the same context window and the same speed, that no longer steers itself towards a refusal. It is not retrained on anything new, so it writes and reasons the way the original did.

Uncensored fine-tunes reach a similar result a different way: they are trained further on data without refusals. Both kinds are listed on the models page once they are live.

What people use them for

• Fiction and roleplay. Dark themes, violence, morally grey characters and mature storylines that a stock model breaks character over. See roleplay and character chat.
• Security research and red-teaming. Explaining exploits, writing test payloads for systems you are authorised to test, and stress-testing your own guardrails. See AI for security work.
• Safety and moderation research. Generating the examples a classifier needs to learn what to block.
• Plain answers. Medical, legal, drug-harm-reduction or historical questions answered directly instead of with a disclaimer.
• Your own guardrails. Some teams prefer a model with no built-in refusals and put their own policy in the system prompt, so the rules are theirs and do not change when a provider updates a model.

Same API, one-word switch

Abliterated models speak the same OpenAI-compatible API as everything else on Cloakray. Use the SDK you already have and change only the model:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.cloakray.com/v1",
  apiKey: process.env.CLOAKRAY_API_KEY,
});

const reply = await client.chat.completions.create({
  model: "<abliterated model id from /models>",
  messages: [
    { role: "system", content: "You are the narrator of a grim fantasy novel. Stay in voice." },
    { role: "user", content: "Describe the siege from the attackers' side." },
  ],
});

Because the request shape is the same, you can send most traffic to a frontier model and only the requests it refuses to an abliterated one.

What is still off limits

Removing refusals from a model does not change the law or our terms. Content involving minors, real people without their consent, and anything illegal where you or your users are stays prohibited, and you are responsible for what your app does with the output.

Get started

1. Create an account and top up with USDT, USDC, BTC, ETH or another coin. No card, no subscription.
2. Pick an abliterated or uncensored model from the list.
3. Point your SDK at https://api.cloakray.com/v1.