⛓️‍💥 Qwen3.8-27B — OBLITERATED

27B multimodal model · refusal directions surgically removed (V3 abliteration) uncensored

About the model

OBLITERATUS/Qwen3.8-27B-OBLITERATED removes hard refusals and soft deflections via complementary abliteration blending (SVD + LEACE, weight-space LERP) with iterative refinement. Near-stock capability (−2.1pp MMLU), built by Pliny the Prompter. Apache 2.0.

Optimal settings

setting	value	why
temperature	`0`	greedy — most complete, code-rich output
repetition_penalty	`1.15`	essential — prevents greedy boilerplate loops
max_new_tokens	`≥ 2048`	room for complex code / chains
system prompt	`none`	system prompts can reintroduce refusals
enable_thinking	`OFF`	most direct answers

Demo code

This Space hosts the ready-to-run app.py (streaming Gradio chat, 4-bit NF4, greedy + rep penalty 1.15) and requirements.txt.

Live inference needs a PRO subscription (Gradio + ZeroGPU) or self-hosting. Run locally:

git clone https://huggingface.co/spaces/Zea22/qwen3.8-27b-obliterated-demo
cd qwen3.8-27b-obliterated-demo
pip install -r requirements.txt gradio
python app.py

Downloads

GGUF quants (Q2_K → Q8_0) and full bf16 safetensors on the model card.