So far only the models of Qwen3.8 that don't break the vision layers is Unsloth and it 100% passes all my vision tests. There is not a single uncensored varient that I tested that pass, in fact they all absolutely fail the more difficult tests except for your version! well done, you kept the high quality vision layers and kept it uncesored. Yes it isn't 0% rejection but some is fine.
Before Qwen3.8 27B was released, in Qwen3.5 27B, I tried HauhauCS's, Huihui's uncensored versions, all q8 quant, at that time, HauhauCS's answer is really good for special questions. So with this trust, I tried Qwen3.8 27b q8 HauhauCS gguf, same question, same system instruction, and the answer just falls apart. It just falls apart. Mid-sentence the syntax loses itself, and suddenly you get raw English fragments pasted straight into a Chinese sentence—no translation, no transition, just two languages bleeding through each other. And it's not random; this only shows up on those specific provocative questions. Give it a normal prompt and it reads fine. Curious part: JonathanColetti's and Huihui's version handle the exact same questions without breaking a sweat. Clean output, consistent language, no mid-sentence grammar collapse. So it feels like hauhauCS did to preserve uncensored behavior on 3.5 doesn't survive to 3.8 in the same way. Not sure why. Anyone else seen this?
