@sam_abdennour — every “unrestricted AI” question is really one question: how far do you want to climb? Each rung below hands the rules to someone else, or takes them back. 
0 a normal website their model, their rules
1 a website YOU pick on their site, your model
2 a community Space their host, open weights
3 your key + your UI your model, your bill
4 a file you own nobody's rules
0 · What almost everyone does — a normal chat site. Refusals are baked in, and the rules change without notice. 
1 · Pick the model yourself — OpenRouter Chat → choose Venice: Uncensored, free. A Dolphin-Mistral build that obeys your system prompt. 1.72 billion tokens ran through it in the last seven days, so the listing is alive. Same site, different brain. 
2 · Let the community host it — Hugging Face Spaces, filtered to uncensored. One Space alone runs Qwen3-235B, Qwen2.5-72B-abliterated and Llama-3.3-70B-abliterated behind a single page. No account, no card, no download. It is somebody’s free GPU though, so a Space can sleep or vanish — open three, keep the one that still wakes up. 
3 · Take the key, keep the interface — Venice.ai for ten free prompts a day, or a Space’s own API, dropped into SillyTavern’s custom-endpoint box. 30 of the 36 providers ModelRanker tracks speak OpenAI-compatible, which is exactly why this rung works with any client you like. 
4 · Own the file — our own drill on running an uncensored 30B on a plain laptop. No host, no account, no revocation. This is the only rung where nobody can change your model on a Tuesday. 
The one honest number — abliteration lowers refusals, it rarely zeroes them. ModelRanker’s measurement: Heretic’s own Gemma test still refused 3 prompts out of 100. Any site promising “answers everything, always” is either lying or about to disappear.
And our earlier sorted list of uncensored chatbot sites is still the fastest place to start if you want the same ladder with prices attached. 
!