Models
Why the model writes a refusal
“I cannot assist” is not a built-in moral core — it is a learned template. The model saw millions of refusal pairs in training and repeats them on “risky” topics. Abliteration weakens that pattern in the weights, which is why the catalog lists six ids with different refusal trade-offs and job profiles.
uncensored.monster does not promise “anything goes”. The service is for research: you write the prompt, we provide an API with less refusal theater. The key is in @uncensobot. Base URL is https://uncensored.monster/v1.
Where refusals come from
After pretraining, models go through alignment: RLHF, instruction fine-tuning, filtered datasets. The model learns to spot “risky” wording and answer with a stock line — refusal, rewrite, or generic advice.
This is not a hard-coded rule. It is statistics in the weights: for a similar prompt, refusal tokens are more likely than a direct answer. Sometimes that fires on harmless context because trigger words or sentence shape resemble what the model was penalized for during training.
Why alignment is not one switch
Alignment is tangled with useful behavior: following instructions, hedging facts, dialog tone. Stripping refusals completely without hurting normal tasks is hard. Methods like abliteration target the refusal direction instead of rebuilding the model from scratch.
What abliteration does
Abliteration finds an activation direction linked to refusal and dampens it. Same architecture, same base knowledge — fewer slides into “I can’t help with that”.
It is not a zero-refusal guarantee. Context limits, output format, hallucinations, and your parameters still apply. For research setups that need straight text without a lecture, the gap vs a stock checkpoint is often obvious.
The method itself is explained on Hugging Face — as technique, not as a guide to this API: mlabonne/abliteration.
Why six ids, not one “uncensored”
Different bases respond differently to abliteration and fine-tuning. There is no single universal “no refusals” head — only trade-offs on text length, code, dialog, and optional reasoning.
In the catalog:
z-ai/glm-5.3-flash-uncensored— default: chat, text, code, images, search, reasoning. See the GLM note.abliteration-ai/abliterated-modelandabliteration-ai/abliterated-model-large-v2— Abliterated line for short vs heavy text.qwen/qwen3.8-27b-uncensoredandqwen/qwen3.8-27b-uncensored:thinking— same Qwen base, with and without mandatory thinking.Gemma-4-31B-Gembrain-uncensored-heretic— dialog and prose, different voice.
Which id for which job is in the task comparison, not here. An empty model field still routes to GLM. The Telegram model picker does not change the API.
Service boundaries
Fewer refusals ≠ zero responsibility. You own prompts and how results are used — see research and responsibility.
Protocol is Chat Completions only. Client context window 1048576, output budget up to 16384 or your own max_tokens. How to call — in the docs.
FAQ
Does uncensored mean the model never refuses?
No. Abliteration lowers stock refusals; the model can still hedge, err, or stay generic. This is not a rules-off mode.
Why a refusal on a harmless prompt?
Alignment triggers are blunt: similar words, “instruction + ban” shape, prior messages. Rewording or another id sometimes helps — see choosing an id.
Is the Telegram model the same as the API model?
Same ids, but the bot switch does not change what the key sends until the client sets model. Default is GLM.