I downloaded an uncensored aggressive Qwen 3.5 model and I can see in its reasoning that it is still limiting responses based on safety guardrails (e.g. violence, NSFW).
Anybody have recommendations for truly uncensored models?
EDIT: I turned off reasoning and I think it’s more uncensored if I’m very specific about what the response should include.
Abliteration techniques might be more limited with reasoning models. I don’t know if they process simply be rehashing the arguments or if there’s more under the hood that would be harder to alter.
I try new models from time to time, including some of the thinking ones, but I’ve always come back to the NeuralDaredevil model, even though it’s “old”. Your results may differ depending on the subject matter, but I can’t think of an instance where I hit a wall. At most, maybe some sidetracking but once I told it to be more open it didn’t hold back.
I’m not sure what the appeal of the thinking mode is. Perhaps on some things it does better, but in watching its reasoning I’ve seen it talk itself out of a good solution too. Which is what you get with typical models when you push the context too far and don’t start a new session, they wander.
Thanks! I’ll check out that model. Is it actually usable or just good at being uncensored?
It’s as good as an 8B can be, but with the right system prompt for your purpose and proper expectations, I think it’s good. I’ve had some other newer 8B that blew up after a few cycles, literally getting stuck on something, but I can’t say this one ever did. But again, even the big models like Claude and the rest work better with short sessions and a specific, detailed prompt to start with. Use a model to make the prompt, telling it to be detailed, concise, and minimize fluff. Less tokens in and out that way, less context drift (hopefully).