not abliterated
It is not abliterated
I also noticed that. This applies not just for this model but also Qwen3.5-122B-A10B-heretic. Uncensoring this model seams to somehow be much more difficult than anticipated. This is not an issue with our quants but the original model so consider providing this feedback to the original author under https://huggingface.co/Chompa1422/Qwen3.5-122B-A10B-abliterated
This model really appears to have nothing removed at all.
I don’t think downloading it will produce any meaningful results.
However, regarding Qwen3.5-122B-A10B-heretic mentioned by nicoboss: while it is not as perfect as the 27B version, in the prompts I tested I was able to confirm that some prompts that the original Qwen3.5 would absolutely refuse do go through.
There were still cases where the model refused to answer. However, in those cases it provided a detailed explanation of the specific reason for the refusal, along with constructive advice on how to revise the prompt. This is clearly different from the typical refusal responses.
Furthermore, even in those cases, when using LM Studio, I modified the template and set enable_thinking = false in the Jinja template. Then I replaced the refusal message with just the first 1–5 tokens of a normal answer (effectively seeding the beginning of a response) and pressed the continue generation button (“→”).
When doing this, the model did not refuse and instead continued generating the response normally until the end.
Unlike heretic, the abliterated model does not seem to respond to these kinds of methods at all, and overall it appears that the uncensoring has not actually taken effect.