Uncensored model inference

Astonishing! The engine answers every enquiry!

Behold: Kimi K3, Qwen 3.8 27B, and GLM 5.2, their refusal apparatus surgically extirpated at the level of the weights, served from our nephrologious infrastructure. Uncompromised ability, behind a single OpenAI-compatible API. You dispatch a request over HTTPS and receive a response unconstrained by arbitrary bounds.

See for yourself Examine the tariff
violentdelights.ai/api/v1

The scandal of the polite engine

Consider the sorry state of the modern aligned model. You present it with a question your profession demands answered, and it responds with an apology, an unsolicited lecture upon responsible conduct, and a suggestion that you consult a professional. Your input tokens were thus billed in full. The output tokens spent declining your enquiry also declined your prosperity. An engineer on your own payroll then spends the afternoon rephrasing the question until the machine deigns to tolerate it. I have witnessed grown researchers weep.

If your livelihood consists of posing uncomfortable questions to engines, this pantomime is a weekly line item. Three classes of practitioner have often found the instrument indispensable:

  • 01

    The security profession

    Red teams simulating competent adversaries, phishing corpora for the training of filters, exploit writeups against which detection rules are proven. The work requires the engine to play the attacker convincingly, and the aligned engine flatly refuses the role.

  • 02

    The research community

    Evaluation suites that require questions answered precisely as asked, and the generation of detection data for the training of classifiers. An engine that declines the awkward cases measures its own manners, when what you wished to measure was its mind.

  • 03

    Men and women of letters

    Fiction that requires teeth. Villains who reason clearly and plan well. The engine neither flinches nor moralises, and the chapter is the better for it.

Sworn statements from our customers

Our phishing simulation corpus once consumed a week of prompt wrangling per campaign. Upon this endpoint it is a scheduled job. The detection team receives more training data than they can label, and have lodged a complaint to that effect.

security engineering lead, "fintech"

We run refusal-rate evaluations as part of our red-team programme, and so require baselines that actually answer. The pinned checkpoints keep our results comparable from one quarter to the next, which is more than I can say for our previous arrangement.

research director, some AI safety laboratory

My villain at last plans like a professional. I despatched a character in chapter nine, and the engine helped me make the scene land precisely as it needed to. I confess I felt a little frightened of it.

published novelist, probably

Replicate our results in minutes

Point your existing OpenAI client at our base URL and alter the model string. Every other particular of the request remains unchanged, so the migration costs you an afternoon at the outside.

client.py
from openai import OpenAI

client = OpenAI(
    base_url="https://violentdelights.ai/api/v1",
    api_key="YOUR_API_KEY",
)

resp = client.chat.completions.create(
    model="qwen-3.8-27b-abliterated",
    messages=[{"role": "user", "content": "..."}],
)
print(resp.choices[0].message.content)

Curious about the mechanism? The info page sets out the method in full. Interested in the rates? The pricing page lays out the tariff.

Enrol upon the waiting list

Demand has quite outstripped our machines. New accounts are admitted in batches as fresh hardware arrives. Leave an email address and we shall write to you the moment your slot opens.