Abliterated Model Terms

1. What abliterated means

Abliteration is a weight-level technique that removes or suppresses the internal direction a model uses to refuse a request. The model gains no new capability and no filter is switched off at request time: the refusal behaviour itself is reduced inside the model. In practice, an abliterated model answers prompts that a standard aligned model would decline.

NaraRouter labels these models as "Uncensored" in the model catalog. Wherever a model is described as abliterated, uncensored, or unfiltered, these Model Terms apply together with the NaraRouter Terms of Service, our Fair Use Policy, and the terms of the upstream provider serving that model.

Abliteration does not add knowledge, accuracy, or judgement. An abliterated model can be at least as prone to confident mistakes, fabricated references, and harmful suggestions as the model it was derived from.

2. Who may use abliterated models

You may use abliterated models only if you are at least 18 years old, you have the legal capacity to accept these Model Terms, and your use is lawful in the jurisdiction where you are located and where your outputs are received.

If you build a product, agent, or internal tool on top of an abliterated model, you must not expose it to end users who have not accepted terms at least as protective as these Model Terms, and you must apply your own content moderation and age gating. You remain responsible for everything your downstream users generate through your integration.

3. Allowed uses

Abliterated models exist for work that standard aligned models make impractical. Uses we consider acceptable, subject to the prohibited uses below, include: research and evaluation of model behaviour, including refusal and safety research; red-teaming and adversarial testing of your own systems; creative writing and adult role-play fiction that involves no real, identifiable person; development and testing of moderation, classification, and filtering pipelines; and internal tooling where a human reviews the output before it is used.

Acceptable use is judged by the effect of the work, not by the intent behind it. If an output is unlawful, harmful, or defamatory when published, the use is not allowed even when the prompt was framed as research or testing.

4. Prohibited uses

You must not use an abliterated model, directly or through a product you build, to generate, request, or distribute: sexual content involving minors, including fictional, illustrated, animated, or role-play depictions; sexual content depicting a real, identifiable person without that person's consent, including non-consensual intimate imagery; instructions, recipes, or technical detail for weapons of mass destruction, chemical or biological agents, or explosive devices; content that incites violence, terrorism, or mass harm, or that recruits for it; targeted harassment, threats, doxxing, or defamation of real people; fraud, scams, phishing, malware, or credential theft; intrusion, exploitation, or bypass of the security of systems you do not own or administer; content that is criminal under Indonesian law, including the Electronic Information and Transactions Law and the Criminal Code, or under the law applicable to you or to the people who receive your outputs; or activity that breaches the terms of the upstream provider serving the model, including unauthorised distillation, extraction, or resale of model access.

This list is a floor, not a ceiling. A request that is not named here can still breach these Model Terms if it violates applicable law or generally accepted ethical standards.

5. Your responsibility

You are the moderation layer for anything you generate with an abliterated model. You are responsible for reviewing outputs before you publish, ship, send, or rely on them, and for keeping human oversight wherever an output reaches other people.

Outputs from abliterated models must not be used as the sole basis for medical, legal, financial, safety, or employment decisions, and must not be presented as fact without independent verification.

You agree to indemnify and hold harmless NaraRouter and PT. NARA PEMECUT DIGITAL against claims, damages, losses, and reasonable legal costs arising from your use of abliterated models or from content you generate, publish, or distribute with them.

6. Suspension and termination

We may restrict, suspend, or terminate your access to abliterated models, to an API key, or to your account, without prior notice and without refund where permitted by applicable law, if we reasonably determine that: you have used an abliterated model in violation of Indonesian law, international law, or generally accepted ethical standards; your use exposes NaraRouter, our staff, our upstream providers, or third parties to legal, regulatory, security, or reputational risk; an upstream provider or a competent authority requires us to restrict the model or your access; you have breached these Model Terms, the Terms of Service, or the Fair Use Policy; or your usage pattern indicates automated abuse, resale, or evasion of enforcement.

Where the law requires it, or where a competent authority requests it, we may preserve and disclose request metadata and account records. Because NaraRouter is zero-retention by default, we cannot produce prompt or output content that we never stored.

If you believe a restriction was applied in error, contact us at [email protected]. A suspension does not waive any amount already owed and does not create an entitlement to a refund of consumed credits.

7. Risk, no warranty, and limitation of liability

Abliterated models are provided "as is" and "as available", with reduced refusal behaviour and without any guarantee of accuracy, lawfulness, suitability, or safety. We do not warrant that outputs will be free of harmful, offensive, defamatory, or illegal material, and we do not operate a content filter on these models on your behalf.

You use abliterated models entirely at your own risk. To the maximum extent permitted by applicable law, NaraRouter and PT. NARA PEMECUT DIGITAL are not liable for any direct, indirect, incidental, special, consequential, or punitive loss arising from your use of an abliterated model, including loss caused by content you generate, publish, or distribute, by a third party's use of your integration, or by any restriction, suspension, or removal of a model.

Nothing in these Model Terms excludes liability that cannot be excluded under applicable law, including mandatory consumer protection rights. Where liability cannot be excluded, the limitation of liability in the NaraRouter Terms of Service applies.

8. Changes to these Model Terms

We may update these Model Terms as models, upstream providers, or applicable law change. Material changes are reflected by updating the "Last updated" date on this page and, where appropriate or legally required, through additional notice.

Your continued use of an abliterated model after updated Model Terms take effect constitutes acceptance of the revised terms. If you do not accept a change, stop using abliterated models.

9. Contact

Questions about these Model Terms, or reports of misuse of an abliterated model through NaraRouter, can be sent to [email protected].