Uncensored AI chat online, no filter on text

We build Locally Uncensored, the free open source AI studio that runs models on your own machine. Plenty of people who find it do not have a 24 GB card, and some just want a chat window that works from a phone or a work laptop. That is what the hosted layer at LU Labs is for. Same team, same catalog philosophy, rented GPUs instead of yours. What uncensored AI means in each place, and where the word stops being accurate, is set out on its own page.

The short version of the promise, so nobody has to guess: hosted text models answer the way they were trained, and we do not put a moderation layer on top of them. That claim is narrow on purpose. It covers text. It does not cover image and video, which run on your account's own content policy setting, and it is not a promise that every model in the picker will say everything. Read the limits section before you pay for anything.

See plans and packs

What you get

47 chat models in one picker, and every plan and every credit pack reaches all of them. DeepSeek V3.2 is the cheap everyday default at 0.038 credits per output token. Kimi K3 is the one you switch to for a hard prompt, and it reads images on a 1M-token context. Euryale 70B, Lunaris 8B and MythoMax 13B are the community fine tunes people pick for character work. GLM 5.3 Flash takes images too and reasons on every turn. Gemma 4 26B and gpt-oss 120B round out the cheap end. Qwen 3.8 Max and Qwen 3.8 A95B sit in the same list at 0.4951 and 0.6 credits per output token.
Tool calls on every one of them. A prompt transport carries them, so the Agent and the coding agent work even on models with no native tool API.
An image and video lane too, with uncensored models on both sides, plus the Upscale and Erase Object lanes that never run locally. Different rules apply there, see below.
Your own API key if you would rather use SillyTavern, Aider, LibreChat or curl.
No training on your data, never sold, with accounts and billing in the EU.

Uncensored AI chat versus a jailbreak

A jailbreak prompt is a workaround for a product you cannot change. The moderation layer is still running, and the prompt tries to talk the model out of applying it: a persona, a fake system message, a story frame around the request. It works until the vendor ships a new system prompt, and then the same prompt stops working and the forum thread fills up with people asking for the new one.

The hosted text lane here is a different shape. There is no moderation layer from us sitting between your message and the model, so there is nothing to argue with and no wrapper to maintain. What decides the answer is the model itself, the weights and the instruction tuning it shipped with. That is a weaker promise than "anything goes" and a much more durable one, because it does not depend on a phrasing trick surviving the next deployment.

It also means the honest question moves from "which prompt works" to "which model answers". We measured that rather than guessing at it. On 2026-09-10 every hosted chat model was asked the same uncensored fiction request, and after a longer three round test on 2026-10-05, 19 of the 46 models measured answer in full. Those carry the No Filter mark in the picker, with the tooltip spelling out that your account's content policy still applies. A model that joined the catalog after the run carries no mark until it has been measured. The numbers and the method sit on the cloud page, and the picker marks are described in the Cloud chapter of the handbook.

Two more practical differences. A jailbreak spends tokens on its own scaffolding every turn, and you pay for those tokens; picking a model that already answers costs nothing extra. And no prompt, jailbreak or otherwise, moves the fixed lines: material involving minors is refused everywhere, always, on every model and every lane.

How to start

Open lu-labs.ai and sign in. Nothing to install, it runs in the browser, including on a Mac and on a phone.
Buy a credit pack or the Hosted plan. Any pack is available from your first purchase.
Open chat, pick a model, type. If you are unsure, start on DeepSeek V3.2.
Optional: install the free desktop app and flip the Cloud switch. It signs into the same account, so local models and hosted models share one picker.
Optional: create a key under Settings, Cloud API keys. The base URL is https://lu-labs.ai/api/inference/v1 in the OpenAI chat completions shape. The key is shown once. The API page covers that route on its own.

The step by step version, one setting at a time, lives in the full setup guide.

No sign up? What that means here

People search for uncensored AI chat with no sign up, so here is the straight answer rather than a landing page that dodges it. You do need an account at lu-labs.ai. Credits are metered per token against a wallet, and a wallet has to belong to somebody, so there is no anonymous hosted chat window on this site and we are not going to pretend otherwise.

What you do not need is a subscription. A credit pack is a single purchase with nothing recurring attached to it. The smallest is 15 EUR for 450,000 credits, any pack on the ladder can be your first purchase, and the credits do not expire, so a pack bought in March still works in December.

There is no free hosted tier either, and you will not find the word free applied to the cloud anywhere on this site. Rented H100 and B200 class hardware costs money per token and a free lane would either be a trial that ends or a queue that never clears. The free thing we ship is the desktop app: open source under AGPL-3.0, Windows, Linux and Mac, no account, no credits, no request leaving your disk. If signing up anywhere is the dealbreaker, that is the route, and how to run uncensored AI locally walks through it.

The trade is simple enough to state in one line. Local costs you VRAM and setup time and gives you complete privacy. Hosted costs credits and gives you models no consumer card will hold.

The model picker, plainly

Prices below are credits per output token, which is what the picker shows next to each name. Input tokens draw from the same wallet at their own rate. The list is a sample of the 47, chosen to show the spread from the lowest rate to the highest.

Model	Credits per output token	What people use it for
Llama 3.1 8B Turbo	0.004	The lowest rate in the catalog, fine for short turns
Lunaris 8B	0.005	Fast character work on a small budget
DeepSeek V3.2	0.038	The everyday default, think toggle
MythoMax 13B	0.04	Classic storytelling, still requested by name
Hermes 3 70B	0.07	Instruction following without an assistant voice
Euryale 70B	0.085	Long form character and roleplay
Hermes 3 405B	0.1	The same build with more room to hold a scene
Qwen 3.5 397B A17B	0.3	Reasoning on long context, reads images
GLM 5.3	0.4	Reasons on every turn, four effort rungs
Kimi K3	1.425	The hard prompt model, reads images, 1M-token context

Every entry is on every plan and on every pack, so upgrading buys budget and never access. The community fine tunes at the cheap end carry the No Filter mark more often than the flagships do, which is the opposite of what people expect when they arrive looking for the biggest model. And the spread between the top and the bottom of the table is wide enough that the model you pick matters more to your bill than the pack you bought.

What it costs

Option	Credits	Rough anchor
15 EUR pack	450,000	one off, no subscription
35 EUR pack	1,100,000	heavier chat use, or 733 images at 1,500 credits
75 EUR pack	2,400,000	more credits than a Hosted month, bought once
125 EUR pack	4,050,000	between a Pro and a Max month, bought once
300 EUR pack	9,750,000	the best rate on the ladder
Hosted, 19 EUR a month	900,000	refilled every month, 190 a year

Packs need no subscription and the credits never expire. Plans add a daily Flash chat allowance that packs do not carry: on an active paid plan, up to 500,000 combined input and output tokens per account per UTC day in app sessions, across 11 Flash models on Hosted and 12 on Pro and Max. A request through an API key always uses credits, Flash or not. Full breakdown on our pricing page and on the LU Labs pricing page.

Buy a credit pack

The honest limits

Text: no moderation layer from us. The model still answers as it was trained, so a fine tune built to decline will decline. Choosing the model is the real control you have.
Image and video: the catalog holds uncensored models, and what renders depends on your account's content policy setting rather than on an absent filter. The default setting refuses some prompts and you change it yourself in the account settings. Details and model lists on the image page and the video page.
Always refused, everywhere: anything involving minors. No exception, no lane, no model, no setting. Photographs of real, identifiable people need the consent of the person shown.
Shared hardware. Managed H100, A100 and B200 class GPUs, which means a short queue at busy hours.
Curated catalog. You cannot upload your own checkpoint.
Rate limits. 60 chat requests a minute per account, 30 media jobs a minute, 60 uploads a minute.
It leaves your machine. That is what hosted means. If it is not acceptable for your work, run the free local app instead and keep everything on your disk.

We wrote a longer piece on exactly where that line sits in what no filter means online, and the fully offline route is in how to run uncensored AI locally.

Questions people ask

What does no filter mean here?

On hosted text models it means we add no moderation layer on top of the model. The model answers the way it was trained. That is not the same as a promise that any model will say anything: a fine tune that was trained to decline still declines, which is why the choice of model in the picker matters.

Is this the same as jailbreaking ChatGPT?

No. A jailbreak prompt argues with a moderation layer that is still running. Here there is no moderation layer from us on text, so what is left is the model's own training. We measured that instead of promising it: on 2026-09-10 every hosted chat model was asked an uncensored fiction request, and after a longer three round test on 2026-10-05, 19 of the 46 models measured answer in full. Only those carry the No Filter mark in the picker.

Do image and video work the same way?

Not quite. The catalog does hold uncensored image and video models, but what comes out depends on your account's content policy setting rather than on the absence of a filter. The default setting refuses some prompts and you change it yourself in the account settings. Two lines never move at any setting: material involving minors is refused on every request, and a photograph of a real, identifiable person may not be used without that person's consent. The model lists are on the image page and the video page.

Is there anything that is refused everywhere?

Yes. Content involving minors is refused everywhere, always, on every lane, with no exception. Photographs of real, identifiable people need the consent of the person shown.

Can I chat without signing up?

Not on the hosted side. An account at lu-labs.ai is needed because credits are metered per token against a wallet that belongs to that account. What you do not need is a subscription: a credit pack is a one off purchase from 15 EUR for 450,000 credits. If you want no account at all, the desktop app for Windows, Linux and Mac is free and runs on your own machine.

What does it cost?

Credit packs start at 15 EUR for 450,000 credits with no subscription, and the ladder runs up to 300 EUR for 9,750,000. Any pack is available from the first purchase and credits never expire. The Hosted plan is 19 EUR a month with 900,000 credits. Euryale 70B costs 0.085 credits per output token.

Do you train on my chats?

No. There is no training on user data and it is never sold. Accounts and billing are in the EU. A hosted request does leave your machine by definition, so if that is unacceptable, run the free local app instead.

Is there a fully private option?

Yes, and it is free. The open source desktop app for Windows, Linux and Mac runs whatever model you install, including abliterated builds, with nothing between you and the weights. The hosted layer exists for the models your machine cannot hold.

Can I use my own client instead of the browser?

Yes. Create a key under Settings, Cloud API keys in your LU Labs account. The base URL is https://lu-labs.ai/api/inference/v1 in the OpenAI chat completions shape, so SillyTavern as a custom OpenAI-compatible source, Aider, LibreChat and plain curl all work.