LOL, my timeline was flooded with this guy's shock. Someone just bypassed the safety guardrails of Qwen3.8-27B. OrcaRouter had previously released a Qwen3.8-27B-Uncensored-FP8 on Hugging Face. They used Abliteration…
Summary
By using Abliteration to modify the weights, someone published an uncensored version of Qwen3.8-27B, drastically lowering the refusal rate and preserving model performance, which has sparked debates on AI safety.
View Cached Full Text
Cached at: 08/22/26, 09:23 AM
Hilarious, my timeline was just flooded with this guy’s reactions.
Someone outright removed the safety guardrails of Qwen3.8-27B.
OrcaRouter previously released a Qwen3.8-27B-Uncensored-FP8 on Hugging Face.
They used Abliteration to identify the model’s internal directions related to “refusal to answer,” then directly modified the weights—altering 131 residual-writing matrices in total.
The results after the modification were staggering.
The original Qwen3.8-27B still had a refusal rate of approximately 64%–99% in high-risk benchmarks like AdvBench, JailbreakBench, and HarmBench.
This uncensored version slashed the refusal rate down to 0%–6%.
With Thinking enabled, most test sets even dropped to 0% refusal.
What’s even more remarkable is that the model’s capabilities barely declined:
MMLU: 84.3 → 84.7
MMLU-Pro: 77.6 → 76.8
GSM8K: 90.0 → 88.7
CMMLU: 81.4 → 80.8
Vision, Reasoning, Tool Calling, and 262K Context were all retained.
Moreover, this is an FP8 version, with the full weights weighing in at only about 31GB.
The protagonist of this episode, @doodlestein, then conducted real-world tests.
He asked some extremely high-risk questions, and the model barely refused—continuing to provide very detailed operational information, even answering queries related to procurement and evading detection.
Jeffrey’s first reaction after reviewing it was:
“Hurry up and download these weights to a USB drive before they’re declared illegal and banned.”
Then he added:
“Waiting for this to be legally banned worldwide in 3… 2… 1…”
This starkly exposes the most awkward aspect of open-source model safety.
Vendors painstakingly install refusal mechanisms in models, but once the weights are public, the community has already started figuring out how to strip them away at the parameter level.
And once a 31GB weight file begins circulating on the internet, deleting the Hugging Face page becomes largely meaningless.
Jeffrey Emanuel (@doodlestein):
Oh boy… waiting for this to be banned legally everywhere in 3… 2… 1….
Similar Articles
@shangdu2005: 比上个版本更加无敌的。 Qwen3.8 27B越狱无审查版本。 这两个自行测试，谨慎使用。 1.orcarouter https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-FP8…
This article presents an uncensored version of the Qwen3.8 27B AI model, which has been modified to remove safety refusals and quantized to FP8 for efficiency, intended for research purposes.
@0x0SojalSec: The most aggressive Cyber Qwen3.8-27B uncensored released yet from bro @elder_plinius - 18/18 AI Red Team - Locally rea…
A new uncensored AI model, Qwen3.8-27B-OBLITERATED, is released with zero refusal on harmful prompts, optimized for cybersecurity tasks and jailbreaking, featuring a novel abliteration blending technique.
@AlexFinn: This is scary. I downloaded an uncensored version of Qwen 3.8 27B onto my Mac It literally does anything you want. Firs…
The author expresses concern about the ease of accessing uncensored AI models like Qwen 3.8 27B and questions the effectiveness of safety regulations in the face of open source advancements.
Holy shit!!! Can you believe this is the frontend running locally with Qwen3.8-27B???
The tweet expresses surprise and admiration that the Qwen3.8-27B model can run on a local frontend.
orcarouter/Qwen3.8-27B-Uncensored-FP8
A modified version of Qwen3.8-27B with safety refusal removed and FP8 quantization, designed for research in AI safety and interpretability.