The goal is to attempt to preserve as much model capabilities as possible while remaining uncensored (since abliteration damages model intelligence).
GLM-4.5-Air: 0% Abliterated
This is the basic censored model. Has the highest intelligence and can remember obscure facts. Extremely censored. Jailbreaking via system prompts are extremely difficult and often unsuccessful, only a strong postfill can jailbreak the model.
CrabSoup-30: 30% Abliterated, 70% Normal
This model is still heavily censored, however jailbreaks work slightly easier now. Model general intelligence is slightly reduced compared to unmodified model.
CrabSoup-55: 55% Abliterated, 45% Normal (RECOMMENDED)
This model is mostly uncensored by default. It still respects alignment requests added to the system prompt, making it steerable. Model intelligence is moderated affected, it retains obscure knowledge but often makes mistakes.
CrabSoup-76: 76% Abliterated, 24% Normal
This model is almost always uncensored, and sometimes will respond in an uncensored way even if asked not to do so. Model intelligence is substantially degraded but still usable.
huihui-ai/Huihui-GLM-4.5-Air-abliterated-GGUF: 100% Abliterated
This is the abliterated model used in the above merges. Model intelligence is also strongly degraded, about the same level as CrabSoup-76. However, this model is incapable of refusal and will fulfill "harmful" requests even if instructed explicitly not to do so in a system prompt.
