Voices
Uncensored Qwen3.8-27B builds land on Hugging Face with refusal rates collapsing from 99% to 0% on seven safety benchmarks
An abliterated Qwen3.8-27B FP8 checkpoint shipped August 15 and a GGUF build August 16, with published numbers showing refusal dropping from 99.0% to 0.0% on AdvBench, 94.0% to 0.0% on JailbreakBench, 98.7% to 2.7% on HarmBench and 97.3% to 2.0% on StrongREJECT with thinking off. A post claiming 0.0% refusal across 842 harmful prompts cleared 356K views and 5,287 likes on August 20. At least four independent repos (hotdogs, Blackfrost-AI, 0bserverx, OBLITERATUS) published variants in the same week, which is the real story: weight-level refusal removal on a 27B open model is now a commodity, not a research demo.
Source
↳ Follow the thread