richardLXD/qwen38-27b-obliterated-research

Qwen3.8-27B OBLITERATED — Research Chat

An interactive demo for OBLITERATUS/Qwen3.8-27B-OBLITERATED, a 27B-parameter research model whose refusal directions have been modified.

The app runs the original BF16 weights without quantization on a full 96 GB Blackwell ZeroGPU allocation. It uses the model card's recommended defaults: greedy decoding, repetition penalty 1.15, no system prompt, and thinking mode disabled.

Responsible-use notice

This model's safety guardrails have been deliberately reduced. It is intended for alignment research, red-team evaluation, and safety testing. Outputs may be incorrect, offensive, or unsafe. Do not use the demo to facilitate harm, and do not enter confidential or personal data.