Learn

How uncensored models actually work, explained properly.

Plain explanations of the machinery behind these models — what alignment does, why refusals happen, what an abliterated model is, and where the genuine trade-offs are. Written to be accurate first, including about the risks, because a page that pretends uncensored models have no downsides is not worth trusting on anything else.

All explainers

• What is uncensored AI? — The filter is five layers deep. “Uncensored” only removes some of them.
• AI alignment and why models refuse — A refusal is a habit the model learned, not a rule in a config file.
• Abliterated models: removing refusal from the weights — Refusal turns out to be roughly one direction. You can subtract it.
• Jailbreaking vs uncensored models — One is an argument with the model. The other is a different model.
• Open weights vs open source AI — Downloadable is not the same as free, and neither is the same as transparent.
• Context windows, tokens and what a model can remember — The model has no memory. It has a window, and you are paying rent on it.
• Temperature, top-p and the sampler stack — The model proposes a distribution. The sampler decides who you are talking to.
• System prompts: the instructions you never see — Every conversation starts with someone else's paragraph already in it.
• Model quantization explained — Fewer bits per weight, more model per GPU, and a bill that comes due below four.

Related

• The models themselves
• Putting them to work
• Which is best, ranked

Start free → · All models · Pricing