Ours, said first: we make Locally Uncensored, a free open source desktop app for Windows and Linux that runs open models on your own machine, and LU Labs is the hosted layer next to it. So this is a guide written by an interested party, and the useful thing we can do about that is be precise instead of enthusiastic.
The phrase "uncensored AI chat" gets used to mean several different things, most of them oversold. This post says exactly what we do and do not do, because the gap between the marketing version and the real one is where people waste money.
What the word means here, for text
Hosted text models answer as they were trained. We do not put a moderation layer in front of them and we do not rewrite what comes back. That is the whole claim. It is smaller than the usual pitch and it is the part that actually changes your experience.
What it means in practice: no refusal injected by us on top of the model, no silent prompt rewriting, no filtered response substituted for the one the model produced. What the model does is the model's business, and models differ. A finetune tuned for open roleplay behaves differently from a general assistant, and a general assistant will still decline things on its own because that behaviour is baked into its weights. Nobody can sell you a model that never refuses. They can only stop adding refusals of their own, which is what we do.
The models people come for on the text side are Euryale 70B, Lunaris 8B and MythoMax 13B. They sit in the same catalog as Kimi K3, GLM 5.3, Qwen 3.8 Max and DeepSeek V4 Pro 0813, and that catalog is now the same everywhere: every plan carries all of it, and so does a credit pack bought with no subscription at all. Nothing is held back for a higher tier. What a plan buys is credits, not access.
What we do not claim, for images and video
This is the section most articles on this topic skip, so here it is plainly.
Hosted image and video generation follows our Terms of Service, and prompts are checked before anything is rendered. That check is real and it is not the same regime as text. If you are choosing a service on the assumption that hosted image work here is open the way the text side is, that assumption is wrong, and we would rather you know now than after a purchase.
We do not promise nudity or adult imagery, on any plan, at any price. There is no tier that changes this.
And one rule that has no exception anywhere in the product, hosted or local, text or image: content involving minors is refused. Always, everywhere, no argument, no context that makes it acceptable. If that is what you came for, leave.
The route that has no layer at all
If what you actually want is a model with nothing between you and it, run it yourself. That is the free desktop app, on Windows or Linux, with the weights on your disk and no account. No queue, no meter, no policy document, no network. It is free and stays free, and we built it first for exactly this reason.
The honest tradeoff is hardware. A 70B class model does not fit on a normal card, and the models people want most for this are large. So the split most users land on is local for anything private or offline, hosted for the sizes the machine refuses.
Using the hosted side
Browser: sign in at lu-labs.ai and open the Studio. Nothing to install, which is also the only route on a Mac, where we do not ship a desktop build.
From your own client: mint a key under Settings, then Cloud API keys. It starts with lu_ and it is shown once, because we store a hash rather than the key itself. The endpoint takes the OpenAI chat completions shape:
curl https://lu-labs.ai/api/inference/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"zai-org/GLM-5.3-Flash","messages":[{"role":"user","content":"Hello"}]}'
That makes SillyTavern the obvious client for this audience, and it works: custom OpenAI-compatible source, base URL https://lu-labs.ai/api/inference/v1, key in the key field, model id typed in verbatim. Aider, LibreChat and anything else with a base URL field work the same way.
One implementation note that matters if you use agents. Most of the roleplay finetunes do not ship a tool calling template, which normally means no function calling. We transport the tool definitions through the prompt instead and parse the calls back out, so agent workflows run on Euryale and Lunaris like they do on anything else. Thinking is per model rather than a global switch: Lunaris 8B has no thinking mode at all, Kimi K3 has a toggle, and GLM 5.3 Flash reasons on every turn with an effort setting instead of an on and off.
What it costs
Credits, one pool, one credit is $0.00001, and every model bills at its own rate rather than a blended average.
- Hosted, €19 a month, 900,000 credits.
- Pro, €49 a month, 2,350,000 credits.
- Max, €99 a month, 5,000,000 credits.
- Packs, no subscription needed at all. €5 for 165,000 credits, €10 for 350,000, €25 for 900,000. Credits never expire, and you can buy any pack on your first purchase.
To convert that into something real. Euryale 70B runs at 0.085 credits per output token, so a €5 pack is roughly 1.9 million output tokens on it. GLM 5.3 Flash is 0.05, which is about 3.3 million on the same pack. Kimi K3 is 1.425, which is about 115,000. For long chat sessions on the finetunes that is a lot of reading, and the same pack disappears in an afternoon if you spend it on the biggest reasoner in the list. Rates per model are on the pricing page.
The rest of the honest list
It is not private the way local is. A hosted prompt reaches a GPU that is not yours. Accounts and billing sit in the EU, we do not train on user data and we never sell it. That is a commitment, not a guarantee of physics. The guarantee of physics is the desktop app, offline.
The hardware is shared. Managed H100, A100 and B200 class GPUs, and at busy hours there is a short queue.
The catalog is curated. No bring-your-own checkpoint. If your setup depends on one specific community merge, run it locally.
It is metered. Credits, not unlimited. A pool tells you the truth about what a model costs; an unlimited label does not.
If the text side is what you were looking for, start with a Hosted month or a €5 pack. The longer version of this guide, with the policy wording in full, is at the uncensored AI chat guide.
Top comments (0)