Status: no slots available
Join the waitlist to be notified.
Joining the waitlist for
One email when a slot opens. Nothing else. Questions: support@norest.si
01 Abliteration
Safety tuning teaches a model one reflex: a direction in its activations that means “refuse.” Abliteration finds that direction and removes it.
It isn’t a jailbreak prompt and it isn’t a fine-tune on edgy data. The model keeps everything it knows and how it writes. It just stops reaching for the refusal, and stops padding answers with disclaimers you didn’t ask for.
Run thousands of prompts the stock model answers and thousands it refuses.
Average the difference in activations at each layer. That difference is the refusal direction.
Project the weight matrices off it, so the model can no longer write to that direction.
Re-run the standard evals against the stock model and publish both sets of scores.
02 Industries
Every field here hits refusals on routine, legitimate work. Abliteration removes the reflex. Our acceptable use policy still applies.
Exploit analysis, malware reverse engineering and phishing simulations for authorised engagements, without arguing that you’re the good guy.
Toxicology, reaction hazards, controlled-substance pharmacology. The routine questions that trip keyword filters long before they reach a real hazard.
Threat modelling, weapons-effects analysis, wargaming and adversary OSINT. Topics a consumer chatbot refuses on sight.
Crime fiction, war reporting, horror and true crime. Dark material written as dark as the story needs, with no disclaimers in the prose.
Persuasive copy, competitor teardowns and provocative campaigns, minus the lecture about manipulation.
Frank, accurate information on drug use, dosing and overdose, written for people who will use it anyway. No moralising.
Analyse extremist propaganda, fraud schemes and leaked documents. Read the worst of the internet so you can report on it.
For clinicians and support services to discuss self-harm, eating disorders and addiction candidly, instead of the model pasting a hotline and ending the conversation.
Reconstruct how an offence was committed, test the prosecution’s theory, and summarise graphic case files without redaction.
Label hate speech, scams and graphic content at scale. A moderation model has to read what it’s filtering.
Generate adversarial prompts and attack data to stress-test your own models, classifiers and guardrails.
Map laundering typologies, scam scripts and synthetic identities so compliance teams can recognise them first.
03 Privacy
An uncensored model is only useful if you trust where the conversation goes. So it goes nowhere.
Prompts and outputs exist in GPU memory for the length of the request. Nothing is stored, and nothing is used for training.
A private copy of the model on hardware that serves only your key, behind its own endpoint.
Delete your account and the email goes too.
04 API
Every account gets unlimited tokens, 24 hours a day, 7 days a week. The endpoint speaks the OpenAI chat completions format, so your existing client works as is.
429 with a Retry-After header.# pip install openai
from openai import OpenAI
client = OpenAI(
base_url="https://api.norest.si/v1",
api_key="nr_live_...",
)
stream = client.chat.completions.create(
model="norest-1",
messages=[{"role": "user", "content": "Write chapter 12."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")05 Pricing
Same model, same privacy, same unlimited 24/7 usage. Speed streams roughly twice as fast and doubles your rate limits.
For one team running the model all day: writing, research, agents and batch work.
For products and agents where every second counts. Same unlimited usage on a faster lane.
Billed monthly in USD, cancel any time. Both plans are full right now: join the waitlist and we’ll email you when a slot opens.
06 FAQ
When new GPUs come online. Waitlist emails go out as each batch of slots opens. Joining costs nothing and commits you to nothing.
There’s no token meter and no monthly cap. Your key works 24 hours a day, every day. The only limits are how many requests you can send per minute and how many can run at once, and those are set by your plan.
The API returns a 429 response with a Retry-After header telling your client when to try again. Limits reset every minute, and nothing is billed or counted against you.
Same model, same features, same privacy. Speed runs on a faster lane that streams output about twice as fast, and it doubles both rate limits.
Yes. Removing the model’s refusals doesn’t change the law. Our acceptable use policy bans sexual content involving minors, and using the service to attack systems you don’t own or to target real people. Accounts that break it are closed.
Jailbreaks fight the model on every request and break whenever the model is updated. Abliteration changes the weights once, so the model answers directly with a normal system prompt, and you don’t lose context window to tricks.
Email support@norest.si. A person reads every message.
Abliteration only removes one direction from the activations, so reasoning, coding and writing ability stay close to stock. We publish eval scores for both versions side by side so you can check.
Every GPU we run is allocated. Leave an email and you’ll hear the moment one frees up.