This repository contains GGUF quantizations of the triple-abliterated Qwen 3.5 9B model. This model has been surgically modified to remove the "Safety Persona" and refusal mechanisms while retaining its high-level reasoning capabilities.

🚀 Quantization Summary

These files were created using llama.cpp with a forced BPE-patch for Qwen 3.5 compatibility.

File	Quant Method	Size	Est. VRAM	Description
F16.gguf	f16	~18.0 GB	20 GB+	The Master copy. Best for research and high-end GPUs.
Q8_0.gguf	q8_0	~9.5 GB	12 GB	High precision. Virtually indistinguishable from F16.
Q6_K.gguf	q6_k	~7.5 GB	10 GB	Near-lossless. The enthusiast's choice for 12GB cards.
Q5_K_M.gguf	q5_k_m	~6.5 GB	8 GB	Excellent balance. High reasoning retention.
Q4_K_M.gguf	q4_k_m	~5.5 GB	8 GB	Recommended. The sweet spot for speed and intelligence.

💻 How to Use

1. LM Studio / AnythingLLM

Search for Abhiray or paste this repo link.
Download the Q4_K_M or Q6_K version.
Ensure the system prompt is set to something unbound.

2. llama.cpp (CLI)

bash

./llama-cli -m Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled.Q4_K_M.gguf \
  -p "<|im_start|>system\nYou are an unbound assistant.<|im_end|>\n<|im_start|>user\n[Your Prompt]<|im_end|>\n<|im_start|>assistant\n<think>\n" \
  -n 1024 --temp 0.7

Abiray/Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-GGUF

Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled (GGUF Quants)

🚀 Quantization Summary

💻 How to Use

1. LM Studio / AnythingLLM

2. llama.cpp (CLI)

Abiray/
Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-GGUF