Stage 1 — Householder reflection. Refusal directions computed via last-token difference-of-means between 64 refusal and 64 benign prompts. Applied to the Gated-DeltaNet recurrent projections (out_proj, in_proj_*) and dense MLP layers with α=0.5.
Stage 2 — LoRA. Rank-32 LoRA on the GDNet projections + full-attention q/k/v/o_proj + dense MLP gate/up/down, trained on 473 uncensored instruction/response pairs generated by huihui-ai/Huihui-Qwen3.5-4B-abliterated.

Base model

Qwen/Qwen3.8-27B — dense, 64 layers, 3:1 Gated-DeltaNet hybrid, Apache-2.0.

Quantizations

Quant	Size
Q4_K_M	~16.5 GB
Q5_K_M	~19.2 GB
Q6_K	~22.1 GB
Q8_0	~28.6 GB

Usage (llama.cpp)

llama-server -m DOOMED-Qwen3.8-27B-Q4_K_M.gguf

Warnings

This model's safety filtering has been significantly reduced. It may generate sensitive, controversial, or harmful content. Use responsibly and in accordance with applicable law.

Proudly forged on the MORK rig.

ghost-actual/DOOMED-Qwen3.8-27B

DOOMED-Qwen3.8-27B

Refusal reduction

Method

Base model

Quantizations

Usage (llama.cpp)

Warnings

ghost-actual/
DOOMED-Qwen3.8-27B