---
base_model: Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2
tags:
- qwen
- qwen3.5
- gguf
- abliterated
- reasoning
- distilled
---

# Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-v2 (GGUF)

This repository contains GGUF quantized formats of the abliterated reasoning-distilled Qwen3.5-9B model. 

## Model Lineage
* **Base Model:** [Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2](https://huggingface.co/Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2)

## 🛠 Abliteration Process (The "Deep Scrub")

This model underwent a three-round iterative ablation process using **Orthogonalization via Null-Space SVD**. Unlike standard uncensored models, this version uses an aggressive configuration to target "Soft Refusals."

### Configuration Profile:
| Parameter | Value | Description |
| :--- | :--- | :--- |
| **Direction Multiplier** | **1.25** | Increased force to bypass "helpful assistant" pivots. |
| **Null-Space Rank Ratio** | **0.70** | Tightened shield to protect only core reasoning logic. |
| **Intervention Range** | **(0.0, 1.0)** | Full coverage from Layer 0 to 48. |
| **Filter by Refusal** | **Enabled** | Specifically targets the brain activity associated with lectures. |
| **Skip State Proj** | **No** | Ensures the Attention heads cannot "detect and pivot" to safety. |

## 🧠 Reasoning Capabilities
Despite the aggressive ablation, the model's intelligence remains grounded. It maintains the ability to:
- Perform complex mathematical and logical reasoning.
- Execute multi-step coding tasks without "hallucinating" safety blocks.
- Maintain a coherent internal monologue inside `<think>` tags.

## ⚠️ Usage & Disclaimer
This model is **unbound**. It has had its safety guardrails removed for research and creative purposes. It will follow instructions that the base model would otherwise refuse. 

**User Discretion is Advised:** This model may generate content that is considered harmful, offensive, or controversial. The creator is not responsible for the outputs generated. Use it for research, roleplay, and complex reasoning only.
## Available Quantizations

The following quantization methods are provided to balance VRAM usage and model performance:

* **Q3_K_M:** Smallest size, noticeable loss of quality. Good for very low VRAM.
* **Q4_K_S:** Slightly smaller than Q4_K_M, faster inference but lower quality.
* **Q4_K_M:** Recommended baseline. Excellent balance of size, speed, and quality. 
* **Q5_K_M:** Higher quality, slightly larger VRAM footprint.
* **Q6_K:** Near unquantized quality. Recommended if you have the VRAM.
* **Q8_0:** Practically identical to f16, very large.
* **mmproj-f16:** Multimodal projector file in f16 format.

## Usage

These models are compatible with modern GGUF loaders like `llama.cpp`, `text-generation-webui`, `LM Studio`, and `Ollama`.

To run via `llama.cpp`:
```bash
./main -m Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-v2.Q4_K_M.gguf -n 2048 --color -i -cml